User manual
LexVoice User Manual
Record meetings in Obsidian and watch a live outline as you go. Afterward you get notes organized by template and a mind-map overview, and you can save people, tasks, and knowledge points to your vault. Audio files, MP4 screen recordings, and transcripts can be imported and organized too. Transcription and AI come from services you configure yourself.
Chapter 1
1. Product overview#
1.1 Workflow#
The simplest way to use it: click "New recording", then read the notes after the meeting. You can skip everything else.
| Stage | What you can do | See |
|---|---|---|
| Before the meeting | Create a preparation note for the agenda, attendees, and open items. It goes to the AI only after you check "Use for transcription & AI" | Chapter 5 |
| Start recording | Pick a template and audio source, then click "New recording". You can record the microphone, computer audio, or both | Chapter 4 |
| During the meeting | The transcript is written to the note and the outline updates with it. You can enter the meeting name, speaker, and outline in the "Meeting info" box, add notes or ask the AI from the input bar at the bottom of the panel, or type in the note | Chapters 4, 5 |
| After stopping | Saves the recording, organizes the notes, generates the overview, names the file, and writes to today's daily note in the background. You can start the next meeting right away | Chapters 4, 5 |
| Review | Play back the recording, match speaker numbers to names, look up details with "Ask", and reorganize with another template if needed | Chapters 4, 5, 9 |
| Capture | In the "Knowledge" tab, scan a note and confirm people, tasks, learning, and vocabulary group by group. Notes about the same matter are grouped under a meeting topic automatically | Chapter 10 |
| Share | Generate an HTML report, PDF report, or email draft | Chapter 10 |
1.2 Processing and recovery#
- Recording: Set by "Processing" under the panel's "More settings". The default, "Both", transcribes as you record, creates notes from the live transcript after you stop, then transcribes the full recording again in the background. The other options are "Live only" and "Post-meeting only". See Chapter 4.
- Import: Audio goes through the "Post-meeting transcription" service and is then organized into notes. Transcripts are organized directly. For files over a service's limit, see Chapter 7.
- Progress: Click the LexVoice status in the bottom status bar to open Tasks, which shows three stages: transcription, AI organization, and writing the notes.
- Errors: If AI organization fails, the recording and raw transcript are both kept. Continuing only does what is unfinished and does not repeat requests. If the note was changed before finalizing, the plugin stops, and the generated notes can be saved as a new note. Recovery actions are in Tasks. See Chapter 5.
1.3 Costs, data, and system requirements#
- Costs: The plugin is free and needs no account. Cloud transcription and AI organization are billed by the provider. Local models have no platform fees, but you install and maintain them yourself.
- Data: Recordings and notes are stored in your vault. LexVoice has no cloud storage of its own. With cloud services, audio, transcript text, and prompts are sent to the provider you choose. See Appendix C.
- System: Obsidian 1.10.0 or later. Desktop (Windows, macOS, Linux) supports all features. Mobile can only record the microphone; streaming transcription, PDF reports, and some other features are unavailable. See Chapter 13.
- Optional: Writing to daily notes requires Obsidian's Daily notes. Pages such as the learning gallery and task board require the Dataview community plugin with JavaScript queries turned on.
Chapter 2
2. Installation and updates#
2.1 Installation#
From the community plugin directory: Open "Settings → Community plugins", turn off Restricted mode if it is still on, click "Browse", search for LexVoice, click "Install", then "Enable".
Manual installation (official files are published only in the public repository Lynn-x/LexVoice):
- Open https://github.com/Lynn-x/LexVoice/releases/latest and download
main.js,manifest.json, andstyles.cssfrom the same release. - Quit Obsidian. In your vault, under
.obsidian/plugins/, create a folder namedlexvoiceand put the three files in it. - Reopen Obsidian, go to "Settings → Community plugins", turn off Restricted mode if it is still on, and enable LexVoice.
2.2 Updates#
Open "Settings → Update", click "Check for updates", then "Install update". "Check on startup" is on by default and shows a notification when a new version is found. The command palette also has "Check for updates" and "Install available update". After installing, restart Obsidian, or turn LexVoice off and on again on the community plugins page, for the new version to take effect.
- Settings: API keys, save paths, custom prompts, and the queue are all kept. When upgrading by hand, replace only the three release files and keep
data.jsonin the same folder. - Download failure: Local files are not touched; try again later. If a write fails, the previous files are restored automatically. Backups are in the
.lexvoice-update-backupsfolder in the plugin directory. - Damaged files: When there is no new version, "Install update" reinstalls the official files for the current version.
- Startup warning that
main.jsandmanifest.jsonversions do not match: The last update was not fully written. Update again in "Settings → Update", or reinstall by hand and restart Obsidian. - Several devices syncing one vault: Update them all to the same version before resuming unfinished transcription. Otherwise it may start over and be billed twice.
Chapter 3
3. Quick start: your first note in three steps#
3.1 Configure a service#
Recording needs both transcription and AI organization configured. Using one service for both is the easiest.
- Open Obsidian settings and click LexVoice on the left to reach "Home".
- Under "Service" in "Quick setup", pick one. In the Chinese interface SiliconFlow is recommended, marked "(recommended)" after its name.
- Paste your key into "API key". If you have no account yet, click "Sign up" to open the key page.
- Click "Test and enable". The notification shows "OK" or "Failed" with a reason for each item. When done, the overall status on the home page shows "Ready". If it shows "Setup incomplete", go to "Settings → Transcription & AI" to see which item is missing.
To use a local service or one not listed here, click "Add provider" in "Settings → Model services", then choose a service for each task in "Settings → Transcription & AI". See Chapter 8. Cloud services receive your audio and transcript. For confidential or private content, local transcription and local models are recommended. See Appendix C.
3.2 Record a meeting#
- Click the list-tree icon "LexVoice live notes" in the left ribbon to open the panel.
- "Template" defaults to "General notes", which suits most meetings. "Audio" defaults to "Microphone only". For online meetings, choose "Microphone + Windows system audio (recommended)" on Windows; on macOS and Linux, install a virtual audio device first. See Chapter 4.
- Click "New recording" and allow microphone access if the system asks. The plugin creates a note in the notes folder, writes the transcript into it, and updates the "Outline" tab along the way.
- Watch the level bar at the top of the panel. If it moves when you speak, sound is being recorded. If it does not, check your device first. See Chapter 4.
- To finish, click the stop square in the panel, or "Stop, merge, and polish" on the floating button.
Optional during recording:
- Fill in meeting info: Enter the meeting name, speaker, and outline in the "Meeting info" box at the top of the recording note. The file name and the note's level-2 headings follow it, and the content is sent to your configured services. See Chapter 4.
- Type your own notes: Type outside the transcript area of the note. On finalizing, this text goes into "手写笔记" in the source material. It is not sent to the AI when generating notes, overviews, or reports; knowledge scans and AI polish of a whole note do read it. See Chapter 4.
- Add during the meeting: Use the input bar at the bottom of the panel to jot down key points, which go into the notes. See Chapter 5.
3.3 Read the notes#
After stopping, keep Obsidian running. You can start the next recording right away. Click the LexVoice status in the bottom status bar to open Tasks and see progress. If something fails, continue or retry there. See Chapter 5. When finished, you see "LexVoice processing complete" and the note opens automatically:
- The file name is "date · topic", or "date · speaker-meeting name" if you filled in meeting info.
- Below the title is the mind-map overview, which may appear a little after the body. The body's structure comes from the template, and tasks can be checked off.
- "## 原始材料" holds the collapsed raw transcript, playback timeline, and more. Click a time link to play back.
- With Obsidian's Daily notes turned on, the summary and tasks are written to "今日会议概要" in today's daily note.
If you see "The full post-meeting transcript is now available. Existing notes were not overwritten; reorganize the note to update them.", the note will not change on its own. To use the full transcript, right-click the note in the "Notes" tab and choose "Reorganize as". See Chapter 9.
3.4 What's next#
- Look up details: Ask about this note in the "Ask" tab. See Chapter 5.
- Fill in names: Once post-meeting transcription has separated speakers, match numbers to names in the "Speakers" card on the "Outline" tab. See Chapter 4.
- Capture: Click "Scan this note" in the "Knowledge" tab and confirm people, tasks, learning, and vocabulary. See Chapter 10.
- Send to others: Right-click in the "Notes" tab and choose "HTML report", "PDF report", or "Email draft" under "Generate". See Chapter 10.
- Existing recordings: Click "Import audio" in the panel. See Chapter 7.
Chapter 4
4. Recording and meeting notes#
4.1 Start, pause, and stop#
- Start: Click "New recording" in the panel (it shows "Start recording" when an unused preparation note is open), click the microphone in the ribbon or on the floating button, or run the command "Start/stop recording".
- Set the audio source for one recording: Run "Record · Microphone only", "Record · Microphone + virtual audio device", or "Record · Virtual audio device only". Windows desktop also has "Record · Microphone + Windows system audio" and "Record · Windows system audio only".
- Pause: Click the pause button in the panel or on the floating button, or run "Pause/resume recording". Anything said while paused is not transcribed.
- Stop: Click the stop button in the panel, "Stop, merge, and polish" on the floating button, or the ribbon microphone icon, or run "Start/stop recording".
New notes are created in the notes folder, named YYYY-MM-DD HHmm by default. "Open sidebar when recording" is on by default, so the panel opens too. After stopping, the plugin organizes the notes, renames the file to "date · topic", writes to today's daily note, generates the overview, and finally shows "LexVoice processing complete". You can start the next recording while this runs. If organizing fails, the task stays in Tasks and both the transcript and recording are kept.
4.2 Audio sources and devices#
Choose from the panel's "Audio" dropdown, or in "Settings → Advanced → Audio input → Recording source". The default is "Microphone only", and phones and tablets can only use the microphone. Click "Auto setup" to pick a source for the current computer.
| Scenario | Windows desktop | macOS, Linux |
|---|---|---|
| In-person meetings, dictation | Microphone only | Microphone only |
| Online meetings, narrating while watching | Microphone + Windows system audio (recommended) | Microphone + virtual audio device |
| Videos, online courses, podcasts | Windows system audio only | Virtual audio device only |
Windows system audio needs no extra software. A virtual audio device must be installed first: VB-Cable on Windows, BlackHole on macOS, a PulseAudio or PipeWire monitor device on Linux. Click "Set up virtual audio" for the steps (on Windows, select a virtual audio device source first to see it). When installed, click "Detect again", then select the device under "System audio input".
Microphone and computer audio are mixed into one track, so you cannot tell which sentence came from which source. See Chapter 13. If "Microphone" is blank, the system default input is used. If the selected device is unavailable, recording reports an error; select again. Use "Test devices" before recording to check the current source. While recording, the top of the panel shows one level row per input. A bar that does not move means that input is getting no sound.
4.3 When transcription happens#
Choose one of three under "Processing" in the panel's "More settings", or change "Enable live transcription" and "Enable post-meeting transcription" in "Settings → Transcription & AI". At least one must stay on:
| Option | While recording | After stopping |
|---|---|---|
| Live only | Transcribes as you record and updates the live outline | Creates notes from the live transcript |
| Post-meeting only | Records only, no live outline | Transcribes the full recording, then creates notes |
| Both (default) | Transcribes as you record and updates the live outline | Creates notes from the live transcript first, while transcribing the full recording again in the background |
When background transcription finishes, the notes are not replaced automatically. To use the full transcript, reorganize the note.
Segmented services cut segments at the "Live update interval", 5 minutes by default, adjustable from 0.5 to 30 minutes in the panel's "More settings → Update". The change applies from the next recording. Streaming services do not cut segments, so this interval only controls outline updates. See Chapter 8.
4.4 Recording files#
| File | Location and cleanup |
|---|---|
| Full recording | Recordings folder (LexVoice/录音 by default), named lex-<YYYYMMDD-HHmmss>.m4a, never deleted automatically |
| Temporary segments | LexVoice/.cache/segments, cleaned up automatically after successful transcription. To troubleshoot, turn on "Keep temporary audio segments" (off by default) |
| Blank short recordings | Run "Clean up empty short recordings". After you confirm, notes of 10 seconds or less with no usable transcript, and their recordings, are moved to the system trash |
| Accidental recordings | "Discard recordings under 3 seconds" is on by default; shorter recordings are thrown away |
4.5 Note structure#
"Consolidated note layout" is on by default. A finalized note contains, in order, document properties, the title "# date and time · template name", the AI-organized body, and "## 原始材料". The source material holds collapsed blocks for your typed notes, recording details, in-meeting materials, the live outline during recording, the live transcript draft, the playback timeline, the original audio, and the transcript. With the setting off, the note keeps the transcript in time order and appends "## 整合版" at the end.
The mind-map overview comes before the body's first level-2 heading and may appear a little after the body. Click a time link in the transcript to play back that position.
4.6 Continuing and merging#
Right-click a note in the panel's "Notes" list. "Continue recording into this note" starts a new recording; after you stop, old and new content are reorganized together and written back to this note. "Merge with previous recording" combines this note and the closest earlier note into a new note, keeping both source files.
4.7 Daily notes and the floating button#
With Obsidian's Daily notes turned on, organized notes are written to "今日会议概要" in today's daily note and updated in place when reorganized. Change "Add today’s meeting summaries to daily note" (on by default), "Daily note heading", and "Daily note template" in "Settings → Advanced → After completion".
The floating button can be dragged anywhere. When idle it shows "打开最近纪要" and "Start meeting recording". While recording it shows "跳到当前录音笔记的转写位置", pause/resume, "Stop, merge, and polish", and a timer. Turn off "Show floating button" or adjust "Floating button size" in "Settings → Advanced → Floating button".
4.8 Speaker separation (multi-channel microphones)#
With a one-receiver, multi-transmitter microphone, the plugin separates speakers by channel, up to 4 channels. Choose "Auto (recommended)" in "Settings → Advanced → Audio input → Speaker separation". This field only appears on desktop when "Recording source" is "Microphone only".
Before recording, click "Test" beside it and speak into each microphone in turn. If the result is "已分离", you are set. If it is "内容相同", set the receiver output to "Stereo" and retry. If it is "未确认", test again.
During transcription each turn starts with "说话人1:". To fill in real names:
- Open the note and expand the "Speakers" card at the top of the panel's "Outline" tab. If you closed the card, reopen it with "编辑说话人".
- Check the unconfirmed segments. You can click the play button to listen.
- Choose a name under "Assign to attendee", or choose "Enter another name", then click "Confirm assignment".
If you add names after the notes are organized, run "Reorganize current note using speaker names". To correct one turn, use "Correct a speaker turn" on the card.
"Post-meeting only" has no segments and uses the post-meeting service's speaker separation. Streaming transcription does not separate speakers. Numbers from the background full transcription do not match channel numbers, so after reorganizing with the full transcript, check names against the new numbers.
4.9 The meeting info box#
New recording notes start with a "Meeting info" box. All three fields are optional and can be changed during the meeting. Every line in the box starts with >, and pressing Enter adds it. A line without > is not part of the box.
> [!info] Meeting info
> Meeting name: Product weekly
> Speaker: Alex
> Outline:
> 1. Last week's progress
> 2. Next quarter's plan
- The file name becomes "date · speaker-meeting name", or "date · meeting name" if only the name is filled in.
- The note's level-2 headings follow the outline. Discussion outside the outline goes after the outline's sections. Items not covered keep their heading, marked as not discussed.
- The content goes to the configured AI service, and to the transcription service if it accepts reference text. The box is not controlled by the preparation materials switch "Use for transcription & AI".
An empty box is removed on finalizing; filled-in content goes into "手写笔记". Continued recordings, preparation notes, and imports do not have this box, and neither do some specialized templates.
4.10 Your own notes#
Type directly in the note while recording. Text outside the transcript area goes into "手写笔记" in "## 原始材料" on finalizing. It is not sent to the AI when generating notes, overviews, or reports; knowledge scans, AI polish of a whole note, and "Ask" during recording do read it.
- Write above the title and first segment, or after the last segment. Text between transcript segments, and text newly written during a continued recording, is overwritten on finalizing.
- To get content into the notes, use the in-meeting additions at the bottom of the panel. See Chapter 5.
- Do not edit the note between stopping the recording and the notes being written. If you do, you will see "笔记已被修改,未覆盖现有内容。已生成的正文可另存为新笔记". Go to Tasks and click "New note" or "Check again".
4.11 Common situations#
- "Microphone access denied": Click "Save audio only" to end the session, or "Open system settings", then start a new recording once permission is restored.
- "Offline · Recording continues": Recording continues locally and catches up automatically when the connection returns.
- "Almost no audio was detected": Check your device selection.
- Recording on a phone: Keep Obsidian in the foreground. Locking the screen or switching apps may interrupt recording.
Chapter 5
5. Live outline and meeting workbench#
5.1 The live notes panel#
Click "LexVoice live notes" in the left ribbon or run "Open live notes". On desktop it opens in the right sidebar; on mobile it is a full-screen tab. It opens automatically when recording starts, which you can turn off with "Open sidebar when recording" in "Settings → Advanced".
The right side of the title bar has "Notes board", "Tasks" and "Settings". The Tasks icon shows how many tasks need attention or are in progress. Below are four tabs: Outline, Notes, Ask, and Knowledge. After you scroll, a short bar stays at the top with the title, "New recording" and the same tabs.
When idle, choose "Template" (what this recording is organized into) and "Audio" (what to record; locked to microphone only on mobile), then click "New recording", "Import audio", or "Import text". "More settings" contains "Processing" (Live only / Post-meeting only / Both), "Update" (the live update interval), "Style", and "Reasoning" (shows "Unavailable" when the AI service does not support adjusting it; see Chapter 9). These are the same values as the settings page and save immediately.
While recording, the control area shows a stop square, a waveform, elapsed time, a pause/resume button, and a level bar for each input. A bar that does not move means that input is getting no sound.
- "Offline · Recording continues": The outline pauses and catches up automatically when the connection returns. You can also click "Reconnect".
- "AI service temporarily unavailable": Local recording continues.
- "Microphone access denied": Recording has stopped and what was recorded is saved locally. Choose "Save audio only" or "Open system settings".
After stopping, you can start the next recording right away while the previous one is organized in the background. The "Meeting info" box in the recording note is for the meeting name, speaker, and outline. See Chapter 4.
5.2 Preparation materials#
Put the agenda, attendee list, and background into a meeting note, and recording and organizing can refer to them. Click the "New preparation materials" icon on the right of the "Add preparation materials" row. The plugin creates a Markdown note in the notes folder and opens it. It contains the property "与会人" and the sections "会议主题", "会议议程", "待解决事项", and "会前附件", which you can edit directly.
To extract materials, expand the collapsed bar and paste text or images, or click the paper clip "Add attachment" or drag in files, then click the up arrow "Write to note".
- Markdown, TXT, PDF, and PNG, JPG, and WebP images are supported. Export Word documents to PDF first. Limits: 20 MB per file, 8 MB per image, 20 pages per PDF, and 60,000 characters of text.
- The topic goes into "会议主题", the agenda into "会议议程", and the attendee list into the property "与会人", filling blanks only. "待解决事项" is never filled in for you. The original file is saved to Obsidian's attachment location, with a link in "会前附件".
The checkbox "Use for transcription & AI" is off by default and set per note.
| State | Result |
|---|---|
| Unchecked | Only Markdown templates with a topic heading plus an attendee or agenda heading can be read on this computer. Nothing else is sent, and you see "材料尚未发送,请先允许使用本次材料" |
| Checked | Free-form text, PDFs, and images are sent to the current AI organization model for extraction. If the model cannot read images, images and scanned PDF pages fail |
When there is a preparation note that has not been recorded, the main button becomes "Start recording", and the recording and notes go into that note. Manual audio imports and in-vault watched folder imports use this note too; automatic imports from a computer folder do not. You can still write materials during recording, but not while the notes are being processed. The command "Edit preparation materials" opens this area directly.
Once checked, the topic, agenda, and attendees are sent as name references to transcription services that support reference text. The live outline and notes use them to understand what matters. Anything not actually said is left out, and a plan made before the meeting is not written up as a meeting conclusion.
If open items are filled in (or, failing that, list items in the agenda), a "关注事项回应" section follows the summary, giving the meeting's response to each item with a collapsed "原文依据". Items with no supporting text are marked "待核对".
When importing audio, click the icon on the "Preparation" row of the import page, add materials, click "Extract materials", check "会议主题", "与会人", and "关注问题", then click "Save". The content applies to this import only. If there is an unrecorded preparation note, that note takes priority.
5.3 The Outline tab#
While recording, the AI merges newly transcribed content into an outline with timestamps.
- The latest section is marked "Generating", or "Paused" while paused, and the panel scrolls there automatically. Click an earlier section to stop and read it, then click "Back to current" to return.
- Click the timestamp at the start of a section to play back. The title bar shows coverage, such as "8/8 segments covered".
- Click "Refresh" to recalculate now. While generating, click "Stop waiting" to cancel.
- After recording, the outline is saved in the note's source material as the collapsed block "录音中实时大纲(草稿)".
Update frequency follows the "Live update interval": 5 minutes by default, range 0.5–30 minutes, shared by transcription and the outline. For most meetings, 3–5 minutes is a good choice. Shorter intervals use more AI. Change it in "More settings → Update" or "Settings → Transcription & AI → Live transcription". A change made during recording applies from the next recording. The outline can use its own "Fast model", chosen in "Settings → Transcription & AI". To turn off the live outline, set "More settings → Processing" to "Post-meeting only".
Open a note when nothing is recording and this tab becomes a review view: a player (play, drag the progress bar, mute, open the audio file), "Speakers" (shown when speaker information exists; see Chapter 4), the outline, and the "Playback timeline" (click a section or timestamp to jump there and play). The icon in the outline's title bar opens or updates the semantic Canvas. See Chapter 10.
5.4 In-meeting additions and instant questions#
While recording with the Outline tab open, an input bar appears at the bottom with the placeholder "Take a note · #concept ?question !highlight @assignee /task". Enter sends, Shift+Enter starts a new line. Content is inserted into the timeline with the recording time. Click × to remove it.
| Prefix | Effect | Calls AI |
|---|---|---|
| None | Ordinary addition, recorded on the timeline | No |
#concept |
Explains what this concept means in the current context | Yes |
?question |
Answers using the current outline and nearby transcript | Yes |
!highlight |
Marks a key point and says why it is worth keeping | Yes |
@name task |
Records an owner | No |
/task |
Records a task; you can add @name inside it |
No |
Next to the input bar are camera and attachment buttons, and mobile adds a photo library button. Files are saved to "LexVoice meeting materials folder", LexVoice/会议资料/ by default, which you can change in "Settings → Advanced → Files and naming". Attachments can be PPT, PDF, Keynote, Pages, Markdown, TXT, and images. When organizing, the AI only gets the file name and link.
When organizing notes, text additions serve as supporting material, and the transcript wins any conflict. /task entries go into the notes' task section, and @assignee entries serve as evidence of ownership. AI replies are not evidence; they are kept in the "会中补充材料" collapsed block in the source material.
5.5 The Ask and Knowledge tabs#
Ask: When nothing is recording, it asks about the open note; during recording or just after, about this recording's note. The AI answers from the raw transcript and the note body, and says "纪要中没有足够依据" when it finds no support. You cannot ask when the note has fewer than 40 usable characters.
- Click a question under "Suggested questions" or "Follow-up questions" to ask it directly.
- Click "Add to note" to the right of an answer to append it to the note's "## 问一问" section. Click "Select multiple", then "Add selected" to write several at once.
- Question history is cleared when you restart Obsidian or close the panel. Only what you wrote to the note is kept.
Knowledge: Open a note and click "Scan this note". The AI finds candidates in four groups, people, tasks, learning, and vocabulary. Confirm them group by group and they are written to the matching library. See Chapter 10.
5.6 The Notes tab#
Lists LexVoice notes in the notes folder. At the top you can search with "Search notes or keywords", switch between "By date" and "By folder", filter by template, or use "Clear filter".
Click a row to open the note. Hovering shows "Rename", and "Retry transcription" if segments failed. The right-click menu has:
For a note that is still processing or failed, right-click it and choose "View task" to jump to its row in Tasks.
- "Move to group". "Continue recording into this note" and "Merge with previous recording"; see Chapter 4.
- "Generate clean transcript" and "Generate overview". "Generate": Email draft, HTML report, PDF report; see Chapter 10.
- "Reorganize as": Check a style in the upper half, then choose a template in the lower half to start. See Chapter 9.
- "Delete transcription record": Asks for confirmation, and you can also choose to delete the recording file.
Derived versions such as clean transcripts sit under their original. Right-click for "Open original", "Regenerate clean transcript", "Generate overview", and "Delete this version". Deleting does not affect the original or the verbatim transcript.
5.7 The notes board#
Click the board icon in the panel's title bar or run "Open notes board" to view all notes as cards in the main editing area. You can set "Group" ("Folder" or "Type"), "Layout" ("Auto", "Single row", "Two rows"), and the "Semantic map" toggle, use "Search notes" and "New group", or click "Meeting topics" to open the meeting topics page. When grouped by folder, drag cards between groups, or use "Move to group" in the card menu. The group menu has "Create or update topic index".
5.8 Tasks#
After recording or importing audio, processing runs in the background; just keep Obsidian running. Click the bottom status bar or "Tasks" in the title bar, or run "Open tasks". The sidebar switches to Tasks, with "Back" at the top.
| Group | Contents |
|---|---|
| Needs attention | Failed or stopped tasks |
| In progress | Queued and running tasks |
| History | Finished and abandoned tasks |
- The right side of each row has the most common action, such as "Stop", "Continue", "Retry", "Open settings" or "Restore". Hover over the button to see whether it is billed.
- Open a row to see the transcribed minutes, the error and more actions: "Generate note" (uses only what was transcribed), "New task" (processes again with the current settings and is billed again), "New note" (saves the generated notes as a new note when the note was changed), "Check again" and "Relink" (after the note was changed or moved), "Locate audio", "Open note", "Copy diagnostics" and "Abandon".
- With two or more tasks in "Needs attention", use "Retry all" or "Abandon all". Abandoning deletes nothing; click "Restore" in "History" to bring a task back.
- "Clear" in "History" removes only the records, not notes or recordings. Records with unsaved results are not cleared.
Chapter 6
6. Long recordings, large files, and recovery#
6.1 Saving while recording#
Do not force-quit Obsidian during recording. With "Post-meeting only" or a streaming service, the whole recording is written to disk only when you stop. If you see "The full recording is temporarily held in memory. Keep the app open, free disk space, then retry.", do not quit Obsidian. Free up space, then retry or continue in Tasks.
6.2 Importing large files#
Handle large files on desktop. Mobile reads the entire file into memory.
| File | Handling |
|---|---|
| Ordinary audio, up to 1 GB | Read directly |
| Audio-only WebM | Imported normally |
| MP4 video | Audio track only. If it cannot be extracted, files up to 1 GB are read whole |
| WAV over 1 GB | If not 16 kHz mono, first converted to 16 kHz mono in segments |
| Other files over 1 GB | Not read; export the audio first |
| WebM video, MP4 with no audio track, fragmented MP4 video | Not read; export the audio first (MP3 or M4A) |
| Ogg, FLAC, and similar | If they do not fit in this computer's memory, you are told to convert to MP3 or M4A first |
For each transcription service's limits, see Chapter 7.
6.3 Empty results#
- Speech present but no text: SiliconFlow finishes the remaining parts and then stops, listing the gaps by segment number and time. Xiaomi MiMo stops, and the audio is kept.
- No recognizable speech in any part: The task stops with no retry offered. Play back the audio and check your recording device.
6.4 Resuming after an interruption#
Click the LexVoice status in the status bar to open Tasks. See Chapter 5.
| Button | Result |
|---|---|
| "Stop" | Sends no new requests; requests already sent finish and their results are kept |
| "Continue" | Picks up from the saved point without repeating finished parts |
| "Retry" | Retries the stage that failed |
| "Generate note" | Makes notes from what was transcribed when the transcript has gaps |
| "New note" | When the note was changed, saves the generated text as a new note |
| "Locate audio" | When audio is missing, opens where it used to be; restore the file, then continue |
After Obsidian restarts, unfinished meetings do not resume on their own; continue them by hand in Tasks. Retry and resume keep the address, model, and speaker options from when the task started, and use the API key of the current connection. A request that timed out with no response may be billed again on retry.
Chapter 7
7. Importing audio and the watched folder#
7.1 Ways to import#
- The "Import audio" icon in the panel, or the command "Import audio for transcription and polishing": The sidebar opens the import page, which lists audio in the top level of the recordings folder. Check files, then click "Transcribe".
- Right-click audio in the file explorer and choose "LexVoice: transcribe and organize". Audio elsewhere is imported this way.
- Select several audio files, right-click, choose "LexVoice: combine N audio segments…", then choose a template.
- Transcripts: The "Import text" icon in the panel, or the command "Import text / Markdown for structured notes". Multiple files are merged into one note in file-name order.
The import page lets you set this import's "Template", "Separate speakers", "Expected speakers", "Main speakers" and "Preparation". Supported formats: webm, mp3, m4a, aac, acc, wav, ogg, flac, mp4, mpeg, mpga, oga. Export audio from other video formats first. For large files, see Chapter 6. Zero-byte files are not transcribed; they are usually cloud placeholder files, so download them fully and try again.
7.2 Transcription services and limits#
Choose the service in "Settings → Transcription & AI → Post-meeting transcription". Post-meeting transcription of recordings uses the same setting. When a file goes over a limit:
| Service | Handling |
|---|---|
| Alibaba Cloud Bailian | With speaker separation on and over 2 hours, split into parts of up to 2 hours each. Check speakers across parts |
| SiliconFlow | The official address is treated as 1 hour and 50 MB per file. With speaker separation off, files are split automatically. With it on, the whole file is submitted first; if rejected, it falls back to a few large parts, 12 at most |
| Xiaomi MiMo | Split into small chunks locally for transcription. A single file can be up to 8 hours; split longer ones first |
The picker's "File exceeds 25 MB; many transcription APIs may reject it. Consider lowering the bitrate." is only a reminder and does not block the import. The table above is what counts.
7.3 Speaker recognition#
Transcription services can separate speakers directly, with no multi-channel hardware. The switch is in "Settings → Transcription & AI → Post-meeting transcription → Separate speakers" and is on by default. Notes are first generated with numbers such as "说话人1", and one person may get several numbers. Fill in names in the "Speakers" card on the panel's "Outline" tab (steps in Chapter 4), then run "Reorganize current note using speaker names".
If an existing note has no speaker numbers, choose "LexVoice: add speaker separation" from the file's right-click menu. After you confirm, the linked recording is transcribed again, which sends one request to the service.
7.4 Failures and retries#
If an import cannot start, you see "Import incomplete:" and the specific reason, such as another meeting still processing or an incomplete service configuration. If transcription fails partway, the task stops in Tasks with its progress kept, and a retry picks up from there. To change the configuration, click "Open settings". For empty results and recovery, see Chapter 6.
7.5 The watched folder#
New audio that lands in a chosen folder is transcribed and organized into notes automatically. It can be a folder in your vault or a sync folder on your computer (Nutstore, iCloud, and so on). Settings are in "Settings → Advanced → Automatic audio import":
| Field | Description |
|---|---|
| Watched folder | Enter a vault-relative path, or click "Select" to pick a computer folder. Blank means not in use |
| Process new files automatically | On by default. Keep it on for manual scans of in-vault folders |
| Archive subfolder | processed by default; applies to in-vault folders only |
| Wait before processing (ms) | 3000 by default, 3000–10000 recommended, so cloud sync can finish first |
| Scan watched folder now | Click "Scan" to process unarchived audio. The command "Scan watched folder" does the same |
| In-vault folder | Computer folder (desktop only) | |
|---|---|---|
| Notes template | Default note template | General notes |
| Source file | Moved to the archive subfolder when done | Stays in place |
| Audio in the note | Kept for playback | Audio is not copied; only the source file name is recorded |
| Failure | Shows "Automatic import incomplete" | Retries automatically after 1, 5, and 15 minutes, 3 times in all. A manual scan can retry again |
7.6 The status bar#
Click the LexVoice status in the status bar to open Tasks. Leaving Tasks does not stop anything, but keep Obsidian running.
| Shows | Meaning |
|---|---|
| Recording 12:30 / Recording paused 12:30 | Recording or paused |
| N pending | Some tasks failed or stopped |
| Task name · progress | A task is running |
| N queued | Tasks are waiting in the queue |
| LexVoice 就绪 | Idle |
Chapter 8
8. Transcription service configuration#
8.1 The three tasks and quick setup#
| Task | What it does |
|---|---|
| Live transcription | Transcribes as you record, for the live outline |
| Post-meeting transcription | Transcribes the full recording again after recording ends and can separate speakers. Imported audio uses this too |
| AI organization | Turns the transcript into notes, reports, and other AI content |
"Settings → Model services" manages services (API key, address, billing). "Settings → Transcription & AI" chooses the service and model for each task.
The quickest way to configure is "Settings → Home → Quick setup":
- Pick one under "Service". SiliconFlow is recommended in the Chinese interface, and Alibaba Cloud Bailian in the English and Japanese interfaces.
- Paste your API key. If you have no account, click "Sign up".
- Click "Test and enable". Once it passes, all three tasks are bound to this service. Tasks it cannot do keep their existing binding.
8.2 Adding and testing connections#
Click "Add provider" in "Settings → Model services". You can add several connections for one service, such as one pay-as-you-go and one plan.
| Group | Services |
|---|---|
| Transcription & AI | SiliconFlow, Alibaba Cloud Bailian, OpenAI, Xiaomi MiMo |
| AI organization only | Poe, OpenRouter, Moonshot / Kimi, DeepSeek, Zhipu GLM, Volcengine Ark, Tencent Hunyuan, MiniMax, StepFun, Tencent Cloud TokenHub, Baidu Qianfan, Huawei Cloud MaaS, iFLYTEK MaaS, JD Cloud, Google Gemini (OpenAI compatible), xAI, Groq, Mistral, Perplexity, StreamLake, Atlas Cloud |
| Local | Local transcription service, Local Ollama, Local LM Studio, Local OpenAI-compatible service |
| Custom | Custom (OpenAI compatible), for relays, enterprise gateways, or self-hosted services |
Click a row to open the "Model providers" dialog. Changes save on the spot:
- Billing, Region: Shown only when the service has several options.
- API key: Click "Get API key" to open the key page.
- Base URL: Change this line for a proxy or gateway. Leave it blank to return to the official address.
- Test connection: Click "Test" to check the address and API key. To verify a model, use "Connection test" on the "Transcription & AI" page.
- Capabilities: Custom connections declare here what they can do; the default is AI organization. They appear in transcription dropdowns only after declaring transcription.
The status on the right of a row is the result of the latest test or call. "Setup incomplete" means a task bound to it is still missing something; hover over the label to see what. If a test result says "API key not verified", test in "Transcription & AI".
- Removing a connection: Click the trash icon. If tasks or unfinished meetings still use it, replace or finish those first. Existing notes and recordings are not affected.
- Address requirements: Public services must use HTTPS or WSS. AI organization only supports OpenAI-compatible addresses, not Anthropic-compatible ones.
- Key security: Keys are obfuscated in the plugin settings file inside the vault. They are not encrypted, and they are not uploaded to LexVoice. If you sync or share the vault folder, the keys go with it.
8.3 Billing and dedicated keys#
Plans and pay-as-you-go use different addresses, and most plans also need a dedicated key. Mixing them gets rejected by the server. In the dialog, choose "Billing" first, then "Region".
| Situation | Services |
|---|---|
| Key has a dedicated prefix | Alibaba Cloud Bailian sk-sp- (shared by Token Plan and Coding Plan); Xiaomi MiMo tp- or ttp-; Tencent Cloud TokenHub sk-sp- (Coding Plan), sk-tp- (Token Plan personal edition); MiniMax sk-cp-; JD Cloud pk- |
| Needs a dedicated key, prefix not documented | Kimi Code, Baidu Qianfan Token Plan, iFLYTEK MaaS plans, Atlas Cloud Coding Plan, Token Plan in Tencent Cloud TokenHub's international region |
| Shares the key with pay-as-you-go, billing set by the address | Zhipu GLM, Volcengine Ark, StepFun, Huawei Cloud MaaS |
Plans can generally do AI organization only; Xiaomi MiMo's Token Plan can also do live transcription. A wrong choice shows "Token Plan does not support this task. Use a pay-as-you-go connection." or the equivalent for Coding Plan.
For Alibaba Cloud Bailian, only pay-as-you-go in Beijing and Singapore supports transcription. Pay-as-you-go in Hong Kong and the US, and all plans, can only do AI organization. Quick setup treats sk-sp- keys as Token Plan; to use Coding Plan, add it in "Model services".
8.4 Choosing a service and model for each task#
In "Settings → Transcription & AI":
- API provider: Points all three tasks at one service in one go. Tasks it cannot do stay as they are.
- Live transcription: The "Enable live transcription" switch, "Provider and model", and "Live update interval (minutes)" (0.5 to 30 minutes, shared by transcription and the outline).
- Post-meeting transcription: The "Enable post-meeting transcription" switch, the "Separate speakers" switch, "Provider and model", and "Expected speaker count" when the service supports it. With it off, notes are created from the live transcript.
- Connection test: At the end of each row, runs a real test with the selected model. Transcription rows send a segment of silence and do not evaluate accuracy.
| Task | Available services and default models |
|---|---|
| Live transcription | SiliconFlow FunAudioLLM/SenseVoiceSmall, Alibaba Cloud Bailian paraformer-realtime-v2 (fun-asr-realtime for Singapore), Xiaomi MiMo mimo-v2.5-asr, OpenAI gpt-4o-transcribe, Local transcription service |
| Post-meeting transcription (with speaker separation) | SiliconFlow XingChenAGI/XingChenASR-Diarize-V3.0, Alibaba Cloud Bailian fun-asr, OpenAI gpt-4o-transcribe-diarize, local WhisperX |
| AI organization | Every connection with an AI organization address; choose your own model |
- Keep at least one of the two transcription types. The recording panel's "More settings → Processing" switches too: "Live only", "Post-meeting only", "Both".
- After turning off "Separate speakers", switch post-meeting transcription to an ordinary transcription model. Bailian does not need to switch. Speaker separation models cannot be used for live transcription.
- The model list is fetched automatically once the API key is in. If a model is not listed, choose "Enter model name manually…".
- Changes during recording only affect the next session. After changing the key, a retry uses the new key.
8.5 Limits by service#
| Service | Limits |
|---|---|
| SiliconFlow | On the China site, SenseVoiceSmall and the XingChen series are currently free, and Qwen/Qwen3-ASR-1.7B is about ¥0.79 per hour. The console is the authority |
| SiliconFlow | The official address is treated as 1 hour and 50 MB per file, and audio over the limit is split automatically. With "Separate speakers" on, up to 12 parts; check speakers across parts in the sidebar |
| Xiaomi MiMo | Accepts up to 8 hours per file. Pauses between chunks are normal pacing. It cannot separate speakers and has no post-meeting transcription address: when using it, turn off "Enable post-meeting transcription" or use another service for post-meeting transcription |
| Alibaba Cloud Bailian | One key covers all three tasks, but Beijing and Singapore keys are not interchangeable. Live transcription needs an address starting with wss. "Connection test" for post-meeting transcription only verifies upload credentials, not the recognition model |
| Alibaba Cloud Bailian | Ordinary transcription handles up to 12 hours. For speaker separation the official recommendation is no more than 2 hours; beyond that, M4A, MP3, AAC, and WAV are split automatically, and other formats may fail |
| Local services | LexVoice does not download models or start the service. The service must accept audio file uploads and return text. The API key can be left blank, and a single request waits up to 10 minutes. WhisperX responses must include segments[].speaker, or inline labels such as SPEAKER_00 in text |
8.6 Streaming live transcription#
When live transcription is bound to Alibaba Cloud Bailian, text flows continuously into the "实时转写中…" quote block in the note, and "Live update interval (minutes)" only adjusts the outline.
- Desktop only. On mobile you are told when recording starts, and the audio is still kept.
- If the connection drops, it reconnects automatically. That stretch has no live text and is filled in by post-meeting transcription after you stop.
- Failed streaming segments cannot be retried offline. Afterward, use "Reorganize" or re-record that part.
8.7 What the vocabulary list does#
For the vocabulary list, see Chapter 10. Terms and name hotwords (with permission) are sent to OpenAI chunked transcription, local services, and custom services; other services do not receive them. After transcription returns, entries of the form "wrong spelling => correct spelling" in the "易错写法" section are replaced on this computer, which works with every service.
Chapter 9
9. AI organization and the mode system#
9.1 Organization model and fast model#
Choose in "Settings → Transcription & AI → AI organization":
- Provider and model: Generates notes, reports, and other AI content. For adding providers, see Chapter 8.
- Fast model: Used for outlines and overviews. The default is "Same as above". Choose again after changing providers.
AI organization only supports OpenAI-compatible Chat Completions APIs. The address can end at /v1 or be the root address.
9.2 Templates#
Set the default template in "Settings → AI organization → Note templates → Default note template". The recording panel's "Template" dropdown changes the same setting. The import dialog and right-click menu let you choose per use.
| Template | Best for |
|---|---|
| General notes (default) | Most meetings |
| Work meeting | Work meetings of any size |
| Seminar | Academic seminars, topic salons, roundtable forums |
| Interview | External interviews, user research, expert interviews |
| Personal notes | Personal dictation, ideas, retrospectives |
| Learning notes | Bilibili, YouTube, courses, lectures, podcasts |
In "Settings → AI organization → Note generation" you can adjust the structure level (Relaxed, Balanced, Rigorous) and further exploration suggestions (off by default).
The "Meeting info" box at the top of a recording note also affects the notes; see Chapter 4. When "Outline" is filled in, level-2 headings follow it. This content is sent to the AI service and to transcription services that accept prompts.
Text written outside the transcript area during recording goes into the "手写笔记" collapsed block in the source material on finalizing. It is not sent to the AI when generating notes, overviews, or reports; knowledge scans and AI polish of a whole note do read it.
9.3 Reasoning levels#
The recording panel's "More settings → Reasoning" has Fast (reasoning off, faster and cheaper), Default (keeps the service's default behavior; this is the default), and Reasoning (reasoning on, for complex content). This setting affects requests such as Ask, Knowledge, title generation, and AI polish. Note organization, the live outline, the meeting overview, the semantic Canvas, and topic organization always use Fast.
Adjustable on Xiaomi MiMo, SiliconFlow, Alibaba Cloud Bailian, DeepSeek, Volcengine Ark, and Zhipu GLM. On OpenAI, xAI, Google Gemini, and Groq, only when a reasoning model is selected. Other services show "Unavailable", and Token Plan addresses generally cannot adjust it either.
9.4 Reorganizing and AI polish#
Reorganizing creates a new note from the saved raw transcript and leaves the original note alone.
- In the recording panel's notes list, right-click a note and open "Reorganize as". Notes that were never organized show "Organize as".
- Check one style if you want: More detail, Concise, More structured, More natural, Moderate expansion.
- Click the target template to start.
The new note is in the same folder as the original, named like "【General notes · More detail】original file name". Right-click it in the list for "Open original", "Generate overview", and "Delete this version".
"Reorganize current note using speaker names" in the command palette reruns the original template with the speaker names filled in from the sidebar. "AI polish: selection or entire note" organizes the selected text (or the whole note if nothing is selected) with the default template and replaces it in place. The original content is not kept, so be careful with important notes.
9.5 Language and translation#
Translation happens as part of organizing; there is no separate button. Under "Language policy" in "Settings → AI organization → Language and translation", choose Follow Obsidian language (default), Keep original language, Translate to target language, or Target language with key original text. With the last two you can set "Target language", "Preserve proper names" (on by default), and "Additional language requirements". The raw transcript and direct quotes always stay in the original language.
9.6 Custom templates#
In "Settings → AI organization → Note templates", click "Open template library" to enter the "Prompt library":
- Built-in prompts cannot be edited, only "Set as default".
- Click "New custom prompt" and fill in "Prompt name" and "Prompt content". The content must include the {{TRANSCRIPT}} placeholder.
- "Improve prompt with AI" rewrites a draft into a full prompt. Check it, then click "Save as default".
- Deletion cannot be undone.
9.7 Titles and automatic renaming#
After organizing, the AI distills a topic of at most 15 characters and writes the file name as "date · topic", such as "2026-05-20 1430 · Weekly-Q2 planning". If you wrote a "Meeting name", it is used directly. File names you changed by hand are left alone. To turn off automatic renaming: "Settings → Advanced → Notes and display → Append topic to file name", on by default.
9.8 Meeting overview#
The meeting overview is a mind map written after the note title and before the first level-2 heading.
- Generating: "Settings → AI organization → Meeting overview" is on by default. It is generated automatically after the note is saved, using one extra model call. To generate by hand, right-click in the notes list and click "Generate overview". Failure does not affect the notes, and you can retry on its own.
- Content: One topic per section of the notes, read from the note body. When there is too much content, the finest details are left out.
- Editing by hand: The overview is a Mermaid code block in the note, so just edit the text. Hand edits are not overwritten automatically. When adding or removing links, update the numbers in linkStyle at the end of the code block too, or it will not render.
- Export: HTML reports, PDF reports, and the PDF in email drafts all include this diagram.
9.9 Long meetings and recovery#
Long meetings are divided by time into parts and organized in turn, with progress shown like "AI 整理 · 3/8 部分". Each part is saved as soon as it finishes.
- After "Stop", no new requests are sent. Requests already sent cannot be refunded.
- Failed or stopped tasks stay in Tasks (see Chapter 5). "Continue" or "Retry" picks up from the break with the original service and model.
- To change models, use "New task", which may upload all materials again and incur costs.
- After Obsidian closes or restarts, unfinished tasks do not start on their own.
- If you edit the note during organizing, it is not overwritten. Use "Save generated notes as a copy" to keep the result.
Chapter 10
10. Knowledge layer#
Save the people, tasks, knowledge points, and proper names in your notes, view them across meetings, or turn them into reports and emails to send out.
Related settings are in "Settings → Library". Four cards at the top show the counts of people, learning cards, tasks, and vocabulary terms. Click a card to open it.
People, learning cards, tasks, and views default to folders of the same names under LexVoice/资料库. The vocabulary list is LexVoice/资料库/词汇表.md, meeting topics are in the "主题" folder next to the notes folder, and HTML reports are in LexVoice/HTML报告. Change locations in "Settings → Library → Storage and privacy". This only affects content created afterward.
10.1 Knowledge capture (people, tasks, learning, vocabulary)#
Open the sidebar's "Knowledge" tab and click "Scan this note". The AI reads the current note and offers four groups of candidates. Finishing one group moves you to the next automatically. A scan reads at most the first 24,000 characters of the note.
| Group | Actions | Written to |
|---|---|---|
| People | Click "Keep", "Merge", or "Ignore" for each entry, and click a name to fix a misheard one. "Add people (N)" keeps all, "Ignore all" ignores all | People library. Changed names are also replaced in the note body and properties, except single-character names |
| Tasks | All selected by default. Edit the text, click "加责任人", "加时间", or "Add a subtask" to add detail, then click "Add tasks (N)" | The "## 待办" section of today's daily note. Without Obsidian's core Daily notes plugin, one file per task in the tasks folder |
| Learning | All selected by default. Click "Add cards (N)" | Learning cards folder, one file per card |
| Vocabulary | All selected by default. Click a word to edit, then click "Add vocabulary (N)" | Vocabulary list. Changed words are also replaced in the note body |
- Each group shows the first 8 entries by default. Click "N more · Expand" to see all.
- After adding, a toast stays at the bottom for about 5 seconds. Click "撤销" to reverse the writes to people, learning cards, tasks, and vocabulary and the replacements in the note body.
- "Rescan" regenerates all four groups. Unconfirmed candidates are overwritten; items already added are not deleted.
- "Review this group again" puts a finished group back in the list. What was written is not deleted.
- Unprocessed candidates are stored in a hidden comment at the end of the note and survive closing the panel, restarting, or changing devices.
With "Settings → Library → Extract knowledge after transcription" on (off by default), a background scan runs after each organization, and learning cards and tasks are added directly without confirmation. People and vocabulary still need a manual scan. Background scan failures are silent.
10.2 People library#
The people library has one Markdown page per person, recording name, role, common aliases, organization, email, and notes. A page's "相关纪要" requires Obsidian's Bases, and "相关待办" requires the Dataview plugin. The overview file 人员库.base is in the views folder.
Extracting people in bulk: Click "Extract people suggestions" under "Settings → Library → Enrich from previous notes". Each round scans at most 20 notes and uses the current AI organization service. Bulk results and "Merge" in Knowledge both go to the "Confirm people assignments" window. For each entry, choose to create a new person or merge into an existing one, choose "Meeting role", and click "Confirm assignment" to save.
When a person has duplicate pages under the same name, wait for devices to finish syncing, then click "Merge duplicate people" under "Settings → Library → Enrichment and deduplication". Duplicates are merged into the main profile, links across the vault are rewritten, and duplicate pages move to the archive folder.
Whether people data is sent to services is set in "Settings → Library → Storage and privacy → People data policy":
| Option | What is sent |
|---|---|
| Privacy first (default) | No people data |
| Name vocabulary | Names and common aliases, up to 80 entries, sent with transcription and organization requests. The first selection needs permission, which "Revoke permission" withdraws |
| Local enhancement | Names and aliases when the transcription service is a local or LAN address; full people data, up to 60 people, when the AI organization service is a local or LAN address |
10.3 Vocabulary list#
The vocabulary list is one Markdown file with six fixed sections: 人名, 品牌/机构, 项目/产品, 行业术语, 易错写法, 其他专有名词. Entries in 易错写法 use the form "wrong spelling => correct spelling", for example cannary release => canary release.
To add terms: Edit the file directly, click "Add vocabulary (N)" in Knowledge, or click "Extract transcription terms" under "Settings → Library → Enrich from previous notes" for a bulk scan. To let scanned notes join bulk extraction again, click "Clear term scan history" or "Clear people scan history" under "Storage and privacy → Note scan history". For how the vocabulary list affects transcription, see Chapter 8.
10.4 Learning gallery, task board, and library overview#
Open these with the four commands below, or with "Open overview", "Learning cards", "Concepts", and "Tasks" under "Settings → Library → Browse and maintain". These pages need the Dataview plugin with JavaScript queries on; otherwise they only show code blocks.
| Command | Shows |
|---|---|
| Open learning card gallery | All cards in the learning cards folder |
| Open concept gallery | Cards with a concept tag; all cards written by Knowledge have it |
| Open task board | Tasks in the tasks folder, plus task lines written by LexVoice across the vault |
| Open subject overview | All of the above, with "All", "Learning cards", "Concepts", and "Tasks" filters at the top |
Check a task on the board and the matching line in the daily note or file updates too. Click a card to jump to its source. Pages are regenerated every time they open, so do not write in them by hand.
For checking or bulk filtering, use "Detail tables" under "Browse and maintain": "People profiles" opens the people table, "All notes" opens the overview of all notes, and "Create missing views" creates any missing Base views.
10.5 Visual reports (HTML / PDF)#
In the "Notes" tab, right-click a note and choose "Generate" → "HTML report" or "PDF report". An AI organization service must be configured first.
- Seminar template: A fixed layout. Before generating, choose a preset or custom color in "Choose report colors". Set the footer company name in "Footer company name (optional)"; if blank, the company tag in the note is used.
- Other templates: The AI reorganizes the content into a general report. Click "保存为图片" at the top of the page to export a PNG, or an SVG if the browser does not allow it.
- Reports use only the note body and do not send the raw transcript. If the body has fewer than 80 characters, you are told the content is too short.
- If the content, model, and report settings are unchanged, generating again reuses the last result without calling the model. Changing colors does not call the model either.
- PDFs can only be generated on desktop, laid out as one long page. Over 18,000 pixels they switch to A4 pagination automatically.
- Files are saved to "HTML reports folder". This and "Open HTML report after generation" are in "Settings → AI organization → HTML report".
10.6 Email drafts#
In the "Notes" tab, right-click a note and choose "Generate" → "Email draft". The plugin creates an .eml file and opens it in your system's default email client. It sends no email and does not call the AI.
- Recipients: Names such as attendees in the note's properties are matched against the people library, and the emails from their profiles are filled in. If nothing matches, you are told, and the draft is still generated.
- Body: Four sections, summary, decisions, tasks, and follow-up. Sections with no content are left out.
- Attachments: The note's Markdown, an automatically generated PDF of the note, and any report of the same name in the reports folder.
Drafts are saved to LexVoice/邮件草稿, and this location cannot be changed. Automatic PDF generation and opening the email client only work on desktop. Check the recipients, body, and attachments yourself before sending.
10.7 Semantic Canvas and topic index#
A semantic Canvas turns one note into an Obsidian Canvas map, with the core proposition in the center and main threads, key points, and details around it. Open a finished note in the sidebar and click the icon to the right of the "AI outline" title (hover shows "Open or update semantic Canvas"). The outline needs at least 2 nodes, and an AI organization service must be configured first.
The first click generates "note name · 语义图.canvas" directly, in the same folder as the note. Once the map exists, the icon opens a menu:
- Calls AI: "Update entire semantic map", "Update this branch", "Expand further".
- Does not call AI: "Open semantic map", "Adaptive layout", "Expand both sides", "Expand right", "Go to source".
Regenerating keeps the positions and sizes you dragged, plus nodes and links you added. Switching layouts rearranges the nodes.
A topic index groups the notes in one group into several cross-meeting topics and marks how they relate. Choose "Create or update topic index" from the group menu in the "Notes board", or right-click a notes folder in the file list and choose "Create topic index Canvas" or "Open or update topic index". The file is "folder name · 主题索引.canvas" in that folder. If the notes are unchanged, the AI is not called.
10.8 Meeting topics and automatic links#
"Settings → Library → Automatic links and meeting topics" is on by default. It links notes about the same matter automatically and does not move the original files.
New notes have their links extracted during organizing, with no extra requests. Older notes and notes with edited bodies are analyzed in the background, sending the title, date, synopsis, agenda headings, a body excerpt of up to 4,200 characters, and the names and descriptions of existing topics. Audio, paths, and the full verbatim transcript are not sent. Background requests may incur costs and pause during recording. For failures, click "Retry unfinished" in settings.
- Note body and properties: Projects and topics get a wikilink at their first substantive mention, and properties collect the links. Headings, code, tables, and existing links are left alone.
- Topic pages: One page per topic, with "最近讨论", "各次会议中的决定", "会议中提出的未决问题", and "相关会议", each entry with a source. "我的补充" at the end can be edited freely.
- Topic Canvases: One Canvas per topic, with a file name ending in "讨论". Layouts, descriptions, and links you adjusted by hand are kept.
To open: The "Meeting topics" button in the "Notes board", the command "Open meeting topic", or "Open" in settings. The overview page lists "持续主题" and "未归类与待处理".
Topic pages only record discussion: A later opinion does not close an earlier question, two topics do not merge automatically, and tasks are not marked done automatically. If you directly edit an auto-generated part, your edit is kept, that page stops updating automatically, and "保留的手工修改" on the overview page lists it.
- Excluding folders: In "Folders excluded from topic organization", write one path per line. Subfolders are excluded too.
- Excluding one note: Right-click it in the file list and choose "Exclude from automatic topic linking". Choose "Resume automatic topic linking" to restore it.
- After turning the switch off, no new links are added. Links, topic pages, and Canvases already generated are not withdrawn.
Chapter 11
11. Settings reference#
The settings page has 7 tabs: Home, Model services, Transcription & AI, AI organization, Library, Advanced, and Update. AI organization, Library, Advanced, and Update are collapsed by section, and only the first section is expanded the first time you open them.
11.1 Home#
When not configured, it shows "Quick setup": choose a service, enter the "API key", and click "Test and enable". The plugin binds live transcription, post-meeting transcription, and AI organization according to the service's capabilities and tests each one. If you have no key, click "Sign up".
When configured, it shows the overall status, the model and billing label for each task, and four "Readiness" cards. Click a card to jump to that setting. "Not configured / Configured" at the top only changes what is displayed, not the configuration.
11.2 Model services#
Ships with SiliconFlow, Alibaba Cloud Bailian, MiMo, OpenAI, and local services. Click a row to open the "Model providers" dialog; changes save immediately. Click "Add provider" to add a connection. One platform can have several.
Each row's status comes from the latest test or call: Available, Invalid API key, Insufficient balance, Network error, Unavailable, Not tested, Not configured, Setup incomplete.
| Setting | Default | Description |
|---|---|---|
| Billing, Region | Depends on the service | Pay as you go / Token Plan / Coding Plan; China, International, Beijing, Singapore, and so on. The address changes with it |
| API key | Blank | May be left blank for local services |
| Base URL | Official address | Change this line for a proxy or gateway. Leave it blank to return to the official address |
| Capabilities | AI organization for custom connections, Auto for others | Auto / AI organization / Transcription / Transcription and AI |
| Test connection | — | Click "Test" to verify the address and key. Does not run a model |
| Authentication | Bearer Token | In "Advanced settings": Bearer Token / API Key Header / Custom header |
| Per-feature base URL, Per-feature API key | Depends on the service, blank | In "Advanced settings", set the address and key for live transcription, post-meeting transcription, and AI organization separately. Leave blank to use the shared API key |
Click the trash icon to remove a connection. A connection still used by a task cannot be removed. Removal does not affect existing notes and recordings.
11.3 Transcription & AI#
"API provider" at the top of the page switches all three tasks to one service in one go. If the model list cannot be fetched, enter the model name directly. Each item's "Connection test" verifies that the API and model work. Keep at least one of live and post-meeting transcription. Changes apply from the next meeting.
| Task | Setting | Default | Description |
|---|---|---|---|
| Live transcription | Enable live transcription | On | Transcribes during the meeting, for the live outline |
| Provider and model | SiliconFlow · FunAudioLLM/SenseVoiceSmall | ||
| Live update interval (minutes) | 5 | Shared by transcription and the outline, range 0.5–30. For streaming services it only adjusts the outline. Changes during recording apply from the next recording | |
| Post-meeting transcription | Enable post-meeting transcription | On | When on, transcribes the full recording again and creates notes from it; when off, uses the live transcript |
| Separate speakers | On | For post-meeting transcription and imported audio. Notes are created with numbers first, and names are filled in later in the sidebar | |
| Provider and model | SiliconFlow · XingChenAGI/XingChenASR-Diarize-V3.0 | ||
| Expected speaker count | Auto | Range 2–100; shown when the model supports it | |
| AI organization | Provider and model | SiliconFlow, model to be chosen | For notes, reports, and other AI organization content |
| Fast model | Same as above | Used for outlines and overviews |
11.4 AI organization#
| Section | Setting | Default | Description |
|---|---|---|---|
| Top | Meeting overview | On | Generates a Mermaid overview after the note is saved, using one extra model call. Can still be generated by hand when off |
| Note generation | Structure level | Balanced (recommended) | Relaxed (mostly prose) / Balanced (recommended) / Rigorous (nested lists) |
| Further exploration suggestions | Off | Lists AI-suggested directions to verify at the end of General notes, kept apart from meeting decisions | |
| Reorganization preference prompt | Blank | Extra rules that only affect the style options in the right-click "Reorganize as" | |
| Language and translation | Language policy | Follow Obsidian language | Follow Obsidian language / Keep original language / Translate to target language / Target language with key original text. The raw transcript is not translated |
| Target language | 中文 | Shown with the last two policies: 中文, English, 日本語, 한국어, Custom | |
| Preserve proper names, Additional language requirements | On, blank | Shown with the last two policies | |
| HTML report | HTML reports folder | LexVoice/HTML报告 | Changes affect new files only |
| Open HTML report after generation | On | Opens in the system default browser | |
| Footer company name (optional) | Blank | For seminar report footers. If blank, the "公司/" tag in the note is used | |
| Note templates | Default note template | General notes | Used by default for recording, importing audio, and reorganizing, and can be switched at use. "Open template library" manages custom templates |
11.5 Library#
The "Library" section has four cards at the top: People, Learning cards, Tasks, and Vocabulary. Click one to open it.
| Section | Setting | Default | Description |
|---|---|---|---|
| Top | Automatic links and meeting topics | On | Extracts links while organizing notes and writes wikilinks, topic pages, and topic Canvases. For what is sent, see Chapter 14 |
| Meeting topics | — | "Open" enters the topic overview; "Retry unfinished" processes again | |
| Folders excluded from topic organization | Blank | One vault path per line, including subfolders | |
| Library | Extract knowledge after transcription | Off | When on, writes learning cards and tasks automatically. People and vocabulary still need confirmation |
| Enrich from previous notes | — | "Extract people suggestions" and "Extract transcription terms" call the AI organization service; "Pending review" and "Dismissed" show results | |
| Deduplicate people | — | "Merge duplicate people" merges by name, updates note references, and archives duplicate pages | |
| Library overview, Detail tables | — | "Open overview", "Learning cards", "Concepts", "Tasks"; "People profiles", "All notes", "Create missing views" | |
| Storage and privacy | Vocabulary file | LexVoice/资料库/词汇表.md | With SiliconFlow, hotwords are not used for recognition; only misspellings are replaced after transcription |
| People folder, Learning cards folder, Tasks folder, Views folder | See Chapter 14 | One Markdown page per person | |
| Note scan history | — | After "Clear term scan history" or "Clear people scan history", notes can be scanned again, which calls the model again | |
| People data policy | Privacy first | Privacy first: do not send people data / Name hotwords: names/aliases only, with permission / Local enhancement: full people context for local services only. "Revoke permission" withdraws name hotword permission |
11.6 Advanced#
The "Audio input" section has three buttons: "Auto setup" picks a recording source for the current system, "Test devices" checks that inputs work, and "Set up virtual audio" opens the installation guide.
| Section | Setting | Default | Description |
|---|---|---|---|
| Audio input | Recording source | Microphone only | Microphone only / Microphone + Windows system audio (recommended) / Windows system audio only / Microphone + virtual audio device / Virtual audio device only. Windows system audio appears only on Windows desktop |
| Microphone | System default input | If the selected device is unavailable, an error is reported; no other device is substituted | |
| System audio input | Not selected | Required with a virtual audio device | |
| Speaker separation | Auto (recommended) | Auto (recommended) / Close / Separate by channel. Shown on desktop with "Microphone only"; "Test" records 5 seconds to check whether channels are separate | |
| Files and naming | LexVoice recordings folder, LexVoice notes folder | LexVoice/录音, LexVoice/转写纪要 | Changes affect new files only |
| LexVoice meeting materials folder | LexVoice/会议资料 | Creates a subfolder per recording | |
| Note file-name format | YYYY-MM-DD HHmm | Same syntax as Obsidian's Daily notes plugin | |
| After completion | Open notes when complete | On | |
| Add today’s meeting summaries to daily note | On | Requires Obsidian's Daily notes. Creates today's daily note if it does not exist | |
| Daily note heading, Daily note template | 今日会议概要, built-in template | The template supports placeholders such as {{date}}, {{note_link}}, {{summary}}, and {{todos_block}}. "Restore defaults" resets it |
|
| Floating button | Show floating button, Floating button size | On, Large | Always visible and draggable; Large / Medium / Small |
| Recording and transcription | Discard recordings under 3 seconds | On | Accidental recordings are thrown away, neither saved nor transcribed |
| Keep temporary audio segments | Off | For troubleshooting; uses more storage | |
| Notes and display | Consolidated note layout | On | AI-consolidated content at the top, collapsible source segments at the bottom. When off, source segments stay in time order |
| Append topic to file name | On | Writes the file name as "date · topic", with the topic at most 15 characters. If the meeting name in "Meeting info" is filled in, it is used. File names you changed by hand are left alone | |
| Open sidebar when recording | On | ||
| Diagnostics and logs | Local diagnostic logs | On | Stored locally only. The same row has "Copy diagnostic report" |
| Diagnostic logs folder, Clear diagnostic logs | LexVoice/系统/诊断日志 | Changes affect new logs only. "Clear" moves log files to the system trash after confirmation | |
| Automatic audio import | Watched folder | Blank (off) | Enter a vault-relative path, or click "Select" to pick a sync folder on your computer (desktop only). Source audio in a computer folder is neither moved nor deleted |
| Process new files automatically | On | When off, scan by hand with the command | |
| Archive subfolder | processed | Processed files from an in-vault watched folder move here | |
| Wait before processing (ms) | 3000 | Range 0–60000, 3000–10000 recommended, so transcription does not start before cloud sync finishes | |
| Scan watched folder now, Clean up empty short recordings | — | "Scan" processes all unarchived audio. "Scan and clean up" moves entries of 10 seconds or less with no usable transcript, and their recordings, to the system trash | |
| Task retries | Maximum retries | 3 | Range 1–10; beyond that, retry by hand |
| Task queue | — | "Open tasks", "Retry all" |
11.7 Update#
| Setting | Default | Description |
|---|---|---|
| Check on startup | On | At most once every 24 hours |
| Install update | — | "Check for updates", "Install update". Backs up current plugin files and settings before installing |
11.8 Settings in the sidebar#
The top of the live notes panel always shows "Template" and "Audio". Click the "More settings" icon to expand the other four:
| Setting | Default | Description |
|---|---|---|
| Processing | Both | Live only / Post-meeting only / Both, matching "Enable live transcription" and "Enable post-meeting transcription" |
| Update | 5 min | This is "Live update interval (minutes)" |
| Style | None | More detail / Concise / More structured / More natural / Moderate expansion, shared with the right-click "Reorganize as". There is no separate entry on the settings page |
| Reasoning | Default | Fast / Default / Reasoning. Shows "Unavailable" when the service does not support it. There is no separate entry on the settings page |
Chapter 12
12. Troubleshooting#
12.1 Diagnostic report#
When asking for help or sending feedback, first copy the diagnostics and paste them to the developer. For an overall report, use the command "Copy diagnostic report". For a single meeting, open its row in Tasks and click "Copy diagnostics".
Reports are redacted and contain no audio, transcript text, or full prompts.
12.2 Common problems#
| Symptom | Cause | What to do |
|---|---|---|
| Transcription reports 429 / rate limiting | Server-side rate limiting. The plugin waits and retries automatically, and keeps finished parts if it still fails | Retry later in Tasks; it continues from what was saved |
| "The transcription service has returned repeated errors" | The plugin pauses requests for up to 30 minutes. Recording continues, and later segments go to the retry queue | Check the network and service status. A manual retry resumes immediately |
| Message that transcription is slower than recording | Audio awaiting transcription has backed up. Later segments move to the background queue and recording is not interrupted | Wait for background processing. If this happens often, raise "Live update interval (minutes)" or change services |
| Transcription request times out | Network, proxy, or local service | Check the network and proxy. For local services, confirm the service is running |
| A segment returned no text | Segments of 30 seconds or more go to the retry queue, with the placeholder "转写失败(空结果,已进入重试队列)" in the body. Shorter ones are treated as having no content | Wait for the queue to retry. If it keeps coming back empty, check whether the wrong device is selected |
| "No speech detected in this segment" or almost no audio detected for the whole session | The wrong device may be selected | Click "Test devices" under "Audio input" in "Settings → Advanced" |
| Note shows "Transcript has known gaps" | Some segments have sound but no text; the rest is kept | Click "Retry" in Tasks, or open the row and click "Generate note" |
| MiMo transcript ends with "本段较长,末尾可能有少量内容未转完" | A chunk hit the output limit. Text already transcribed is kept | Play back and check this part, or use another transcription service |
| Invalid API key, insufficient balance, or model not found | A configuration or account problem; retrying unchanged gives the same result | Click "Open settings" in Tasks, fix it, then come back and retry |
| Temporary segment or audio not found | The segment cache was cleaned up. "Retry all" does not cover these tasks | Click "Retry" on each in Tasks, and the plugin recovers them from the full recording |
| Original audio missing | The original audio was moved or deleted | Click "Locate audio" in Tasks, restore it, and retry |
| Organization task waiting for configuration | AI organization configuration is missing or the service has a problem | Configure AI organization, then run "Retry all failed tasks" |
| Streaming segment failed | The wss connection failed or dropped. These segments do not enter the retry queue | Recording is not affected. Afterward, use "Reorganize" in the note, or switch to a segmented service |
| "笔记已被修改,未覆盖现有内容" | The note was changed during processing | Click "New note" in Tasks, or "Check again" and continue |
| Checkpoint cannot be read | Saved progress cannot be read | Do not delete files. Send the output of "Copy diagnostics" to the developer |
| Outline does not update for a long time | The outline refreshes with the live update interval | Lower "Live update interval (minutes)" |
| Cannot install or enable the plugin | File location or version problem | Confirm the three files are in .obsidian/plugins/lexvoice/, Obsidian is 1.10.0 or later, and Restricted mode is off |
| "LexVoice 版本错位" | The last update may not have finished writing | Click "Install update" in "Settings → Update", or reinstall by hand from the GitHub Release and restart Obsidian |
| Microphone access denied | The system revoked microphone permission during recording. What was recorded is saved | Click "Open system settings" and start again after granting access, or click "Save audio only" |
| Selected microphone or virtual audio device is unavailable | The selected device cannot be opened | Select again under "Audio input" in "Settings → Advanced" |
| Imported file skipped (0 bytes) | A cloud placeholder file has not finished downloading | Download it fully and import again |
| Files in the watched folder not processed, sync conflict message | The file name carries a conflict marker, common with Nutstore and Dropbox | Resolve the conflict by hand and put it back, or scan by hand |
| "Cannot read audio from this file", file too large, or a format that can only be decoded whole | Desktop does not read WebM video, MP4 with no audio track, fragmented MP4 video, or files over 1 GB (except WAV). Ogg, FLAC, and similar formats of roughly 15 minutes or more do not fit in memory | Export to MP3 or M4A first, then import |
| Message that a single audio file can be at most 8 hours | A limit when transcribing with MiMo | Split into parts of up to 8 hours and import |
| Quick setup says the service returned no model usable for AI organization | The service returned no usable model | Choose a model under AI organization in "Transcription & AI", then test again |
12.3 Tasks#
See Chapter 5 for how to use it. Retryable transcription tasks run once when the plugin starts and switch to manual after reaching "Maximum retries"; network or service outages do not count. Failed or stopped meetings do not start on their own after a restart; continue them in Tasks.
Chapter 13
13. Known limitations and best practices#
13.1 Mobile#
Available: Microphone recording (system default input), segmented or whole-file cloud transcription, AI organization, importing text, browsing notes, the settings page, the live notes panel, and the right-click "Move to group".
- The only recording source is "Microphone only", and devices cannot be selected.
- Streaming transcription is unavailable. If a streaming service is selected, you are told and the audio is kept; switch to a segmented service.
- PDF reports and automatic import from computer folders are desktop only. Handle long audio on desktop too.
- Keep Obsidian in the foreground while recording. Locking the screen or switching apps may interrupt recording.
13.2 Speaker attribution#
| Source | Condition | Reliability |
|---|---|---|
| Multi-channel hardware | A one-receiver, multi-transmitter microphone, with "Speaker separation" test result "已分离" | Most reliable, decided by physical channels |
| Transcription service separation | "Separate speakers" on, for post-meeting transcription and imported audio | Gives numbers such as "说话人 1、2、3". Accuracy depends on the service and recording quality |
| No speaker information | Live transcription only, or "Separate speakers" off | Attribution in the notes can only come from forms of address and context, and may be wrong |
For anything involving responsibility, commitments, or key decisions, play back the original audio and check.
- Match numbers to real names in the sidebar. One person may have different numbers when audio is split; listen in "Unconfirmed segments", then assign.
- If an existing note has no numbers, right-click "LexVoice: add speaker separation". If individual sentences are attributed wrongly, right-click "LexVoice: correct speaker assignment".
13.3 Live update interval#
| Scenario | Recommendation |
|---|---|
| Regular meetings | 3–5 minutes |
| Need frequent outline refreshes (such as negotiations) | 1–3 minutes; LLM calls rise with it |
| Long lectures, only the final notes matter | Choose "Post-meeting only" under "Processing" in the sidebar |
| Transcribing with MiMo | With the interval at 0.5–1 minute and the live outline on, transcription lags a beat. For smooth in-meeting use, choose a segmented service such as SiliconFlow, or a streaming service |
13.4 Other limitations#
- The official SiliconFlow address is treated as 1 hour and 50 MB per file, and is split automatically beyond that.
- When transcribing by channel, overly long segments fall back to mono, so do not set the interval too high.
- Mixed sources are recorded as one track and cannot be split into two afterward.
- Text written between transcript segments is overwritten on finalizing. Text outside the transcript area goes into the collapsed callout "手写笔记" on the first finalizing.
- Ask records are cleared when you close the panel or restart, except answers written back to the note.
- With several devices syncing one vault (including the plugin folder), update them all to the same version before resuming unfinished transcription. Otherwise it may start over and be billed twice.
Chapter 14
14. Appendices#
Appendix A: Commands#
In the command palette (Ctrl/Cmd+P) they start with "LexVoice:".
| Command | What it does |
|---|---|
| Start/stop recording | Starts recording when idle; stops and begins organizing while recording |
| Pause/resume recording | Pauses or resumes the current recording |
| Record · Microphone only / Microphone + Windows system audio / Windows system audio only / Microphone + virtual audio device / Virtual audio device only | Records from this source once without changing settings. The two with Windows system audio appear only on Windows desktop |
| Edit preparation materials / Edit preparation materials for the current note | Opens the preparation materials area in the sidebar or for the current note |
| Show/hide floating controls | Toggles the floating button |
| Import audio for transcription and polishing | Opens the import audio page in the sidebar |
| Import text / Markdown for structured notes | Hands an existing transcript to the AI for organizing |
| Scan watched folder | Processes unprocessed audio |
| AI polish: selection or entire note | Organizes the selection or the whole note |
| Reorganize current note using speaker names | Reorganizes with the current note's mode, using confirmed names |
| Generate overview | Generates or updates the mind-map overview for the current note |
| Generate an HTML report from the current note / Generate a full-page PDF report from the current note | Generates a visual report; PDF is desktop only |
| Open live notes / Open notes board / Open meeting topic | Opens the sidebar, the notes board, or the topic overview |
| Retry unfinished topic organization | Reprocesses unfinished topic links |
| Open learning card gallery / Open concept gallery / Open task board / Open subject overview | Opens the matching aggregate view |
| Scan notes with AI for people suggestions | Scans past notes in bulk and produces people candidates |
| Open tasks | Opens Tasks in the sidebar |
| Retry all failed tasks | Retries everything retryable in the queue |
| Copy diagnostic report | Copies the redacted diagnostic report to the clipboard |
| Clean up empty short recordings / Clean up expired audio segment cache | Cleans up notes and recordings of 10 seconds or less with no usable transcript; cleans up cached audio older than 7 days that the queue does not reference |
| Migrate legacy note properties | Fills in properties for old notes in the notes folder |
| Check for updates / Install available update | Checks for or installs a new version |
Appendix B: Where data is stored#
| Data | Default location |
|---|---|
| Plugin settings (including obfuscated API keys) | .obsidian/plugins/lexvoice/data.json |
| Update backups | .obsidian/plugins/lexvoice/.lexvoice-update-backups/<timestamp>/ |
| Full recordings, transcribed notes | LexVoice/录音, LexVoice/转写纪要 |
| In-meeting materials | LexVoice/会议资料/<recording time>/ |
| Segment audio cache | LexVoice/.cache/segments (cleaned up automatically after 7 days if the queue does not reference it) |
| HTML / PDF reports | LexVoice/HTML报告 |
| Email drafts and attachments | LexVoice/邮件草稿, LexVoice/邮件草稿/附件 |
| Topic pages | The "主题" folder next to the notes folder (LexVoice/主题 by default) |
| Vocabulary list, people, learning cards, tasks | LexVoice/资料库/词汇表.md, LexVoice/资料库/人员, LexVoice/资料库/学习卡片, LexVoice/资料库/待办 |
| Aggregate views, people library | LexVoice/资料库/视图 |
| Duplicate people archive | LexVoice/资料库/归档/重复人员 |
| Diagnostic logs | LexVoice/系统/诊断日志 (daily .jsonl) |
Paths for recordings, notes, in-meeting materials, reports, the library, and diagnostic logs can be changed in settings. Existing files are not moved automatically. The locations of email drafts and the segment cache cannot be changed.
Appendix C: Privacy notes#
| Content | Handling |
|---|---|
| Audio, transcript text, prompts | Sent to the transcription and AI organization services you configure. LexVoice runs no servers and has no ads, behavior analytics, or telemetry |
| Meeting info | The meeting name, speaker, and outline in the "Meeting info" box are sent to the configured services, regardless of the preparation materials switch. Leave them blank and nothing is sent |
| Your own notes | Not sent to the AI when generating notes, overviews, or reports. Knowledge scans and AI polish of a whole note do read them |
| Preparation materials | "Use for transcription & AI" is off by default and set per note. Original attachments and the extraction cache stay local |
| Automatic links and meeting topics | On by default. Organization requests may include existing topic names. Additional analysis sends the title, date, synopsis, agenda headings, a limited body excerpt, and the names and descriptions of existing topics; descriptions may contain past meeting information. No extra audio, paths, or full verbatim transcript are sent |
| People data | "Privacy first" by default, so nothing is sent. "Name vocabulary" sends names and common aliases once permitted, and can be revoked at any time. "Local enhancement" uses full people context only with local or LAN services |
| Other network requests | Checking for updates contacts GitHub and fallback download sources. Diagnostic reports can only leave your computer if you copy them yourself |
| API keys | Stored obfuscated (not encrypted) in data.json inside the vault. Do not share or sync the vault folder with anyone you do not trust |
| Local cleanup | Recordings, notes, and diagnostic logs are moved to the system trash and can be recovered. The segment audio cache in its default location is deleted directly |
Turning off preparation materials or automatic links does not recall requests already sent. For sensitive content such as client data, medical, legal, or HR matters, use local transcription and local models, and get the consent of those involved before recording.
This manual covers LexVoice 3.0.0. The settings page of your installed version is the authority on the interface and default values. Report issues: https://github.com/Lynn-x/LexVoice