English | 简体中文 · Release notes · 中文文档站
LexVoice is an Obsidian plugin for recording audio, transcribing speech, building a live outline while you record, and turning meetings into reusable Markdown — todos, learning cards, people records, and ASR hotwords.
It is not a hosted cloud service and ships no API keys. You connect your own speech-to-text (ASR) service and, optionally, your own large language model (LLM). Recordings stay in your vault; nothing is uploaded to any LexVoice server (there is none).
Official downloads and updates: Lynn-x/LexVoice Releases. From 2.2.0 onward, public releases contain installation files and user documentation. Development source remains private. New material uses the LexVoice Proprietary Software License; historical MIT material retains its original license.
LexVoice supports desktop and mobile Obsidian workflows. Mobile recording uses the device microphone and supports segmented or whole-audio transcription after capture. System audio, virtual audio devices, multichannel capture, desktop device diagnostics, and realtime streaming ASR providers that require custom authentication headers require the desktop app.
- English and Japanese: the interface follows your Obsidian language, alongside Chinese. No separate plugin language setting is needed.
- Localized built-in AI templates: built-in note templates and the default AI output language follow Obsidian. Existing explicit output-language preferences and custom prompts are preserved. Raw transcripts retain the spoken language.
- Layout fixes: long menu labels wrap, recording level meters no longer overlap status text, and speaker inputs remain usable in narrow dialogs.
- Existing notes and custom templates are not automatically translated. Unchanged built-in daily-note templates follow the host language.
See the 2.5.0 release notes.
- Automatic links and MOCs: extract evidence-backed relationships during organization, link existing notes and collect related meetings in topic pages without manual classification or guessed identities.
- Persistent topic canvases: add related meetings incrementally, preserving source links and manual layout.
- One meeting interval: the sidebar and settings share a 0.5–30 minute cadence. Upgrades retain the previous outline/segment interval; streaming transcription remains continuous.
- Task panel: keep details, saved results, stop, resume and retry actions with their meeting.
- Long-audio recovery: with XingChen speaker diarization enabled, try whole compressed audio first, then use a bounded coarse fallback when needed. Common single-track AAC/Opus M4A can be split without decoding the whole recording; completed results are reused.
- Speaker attribution: keep chunk-local labels separate from meeting participants, with playback, roster selection and batch assignment for unresolved fragments.
See the 2.4.0 release notes. Automatic linking retains the existing topic setting and may process older notes when enabled. Empty ASR responses remain unfinished, not automatically confirmed as silence.
- Folded source material: newly organized notes keep the minutes visible and put recording information, live drafts, audio and full transcripts in native collapsed Obsidian callouts. Empty draft segments are summarized instead of repeated.
- Readable outline: expand the recording outline to see timestamps and nested points in separate columns. Older HTML outline blocks remain supported.
- Speakers in Outline: the speaker panel returns to Outline, where it can be folded or closed. Confirmed transcripts take precedence over live-draft labels.
- Editing compatibility: folded transcripts remain available for reorganization, retries, speaker naming and turn corrections.
See the 2.3.1 release notes for the earlier update.
The collapsible preparation area at the top of the sidebar accepts pasted text and Markdown, TXT, PDF, PNG, JPEG or WebP attachments. Export Word documents to PDF first. Images and scanned PDFs require an AI organization model with image-input support.
Attendees are written to YAML properties. The topic, agenda and manually entered questions stay in the note body. The live outline and final minutes are written into the same meeting note, preserving preparation content and attachment links. You can edit the preparation directly in Markdown.
Using preparation for transcription and organization is off by default. When enabled, names and background can guide interpretation, and final responses include verifiable transcript excerpts. Items without evidence remain pending review. Important unplanned discussion remains part of the minutes.
Chapters grow as you record, so you can glance at "what was just discussed" mid-meeting instead of waiting until the end. After recording, chapters link to the player — click a chapter to jump to that position in the audio. When recording stops, AI completes the chapters into a full set of meeting notes.
The sidebar and settings share one meeting update interval, from 0.5 to 30 minutes. Segmented transcription and outline updates use this cadence; streaming services keep transcribing continuously. Changes during a recording apply to the next recording.
While recording, jot live notes under the outline. The first character can trigger different handling:
Trigger the AI assistant:
#term— hit an unfamiliar term? Type#<term>and the AI explains it in the context of the current discussion.?question— type?<question>and the AI answers using the current transcript and outline.!highlight— type!<point>to mark something important and have the final notes treat it accordingly.
Mark only (no AI call):
@assignee— record "@alice follows up"; the final notes prefer assigning that todo to them./todo— type/<action>to capture an explicit todo candidate.
Half-width and full-width symbols are both accepted. In-meeting notes are fed into the final summarization prompt as clearly-labeled "live supplementary material", never mixed into the raw transcript.
Ask follow-up questions when the final notes miss a detail or you want to revisit a specific part of the discussion. LexVoice answers from both the organized note and the preserved raw transcript. Useful answers can be written back to one compact Ask this note section in the Markdown file.
In standard meeting and learning-note modes, long recordings are organized in recoverable parts instead of relying on one all-or-nothing LLM response. LexVoice builds a global topic map, saves each completed part as a local checkpoint, and assembles the final note in time order.
If a request is interrupted or a model reaches its output limit, completed work is reused and only unfinished parts are retried. The raw transcript remains available, and an incomplete result is shown as partially completed rather than being saved as an empty note.
With SiliconFlow XingChenASR-Diarize-V3.0 and speaker diarization enabled, LexVoice tries whole compressed audio first. Explicit size/duration rejection or a large-file HTTP 500 can trigger one coarse fallback of at most eight parts, never recursive splitting. This recovery policy does not apply to ordinary transcription with diarization disabled. Common single-track AAC/Opus M4A can be split without full PCM decoding. Completed parts are retained; empty responses remain unfinished. Complex containers may still be unsupported, and provider success is not guaranteed.
The processing panel separates transcription, AI organization, and Markdown writing. It shows the active stage, recent activity, failures, and retry or cancel actions. Failed transcription and failed AI organization remain distinct so you can resume from the step that actually failed.
Details and saved results stay with their current meeting. Recent completions can be collapsed or opened directly. Closing the panel does not stop processing; stop and resume apply to the corresponding meeting.
When post-meeting diarization is enabled, LexVoice keeps provisional speaker labels through transcription so the meeting does not stop for identity confirmation. Afterwards, use the collapsible speaker panel in Outline to select a name from the preparation attendee list, enter a new name, merge labels that refer to the same person, or correct a specific turn. The system does not infer a real identity from the meeting text alone.
Chunk-local labels retain their provenance instead of counting as new meeting participants. Unresolved fragments support playback and batch assignment, without forcing a match based on the expected speaker count.
After each note, AI splits the content into four candidate groups you review assembly-line style — keep / merge / ignore:
- People — adjudicated one by one
- Todos — selected by default; edit owner, due date, sub-tasks
- Learning — concepts, mechanisms, cases, opinions, Q&A
- Hotwords — names, organizations, brands, terms, to improve later ASR accuracy
LexVoice turns reusable meeting content into standalone Obsidian objects — people profiles, todo cards, learning cards, ASR hotwords, and concept / todo / learning-card walls. Everything lives in your own vault; the next time the same person comes up, it links to the existing profile.
Topic memory can connect multiple minutes about the same subject without asking you to sort every new note. Topic pages keep a source link for each summarized item and separate recorded decisions from unanswered questions. It is deliberately conservative: a later discussion does not automatically close an earlier question, merge two topics, or mark a task complete. You can pause the background work, exclude folders, or exclude an individual note in Library settings.
Organization can produce validated relationships for inline wikilinks, the managed lexvoice_links property and MOCs. People only link to unambiguous existing notes. Original material, code, existing links and custom properties are preserved.
Each topic maintains a Canvas with dates, brief contributions and source links for the latest 12 related notes; its MOC provides all sources. Updates preserve manual layout, annotations and connections while handling new meetings and renamed or moved files. Topic pages link to their canvases, which do not make additional model requests.
Edit owner, due date and sub-tasks inline at the candidate stage — no dialogs. Stored todos use standard Markdown task syntax (recognized by plugins like Tasks). Source information is preserved on delete / redo for traceability.
- Level meters before and after recording show whether the mic and system audio are actually working.
- Audio inputs remain user-selectable. Device names are displayed without speculative virtual or remote-device labels.
- A device check in settings diagnoses "recorded but silent" problems.
- Compatible independent multichannel input can be detected and transcribed by channel, with speaker labels that can be mapped to names. Separation stays off when independent channels cannot be verified.
- Deleting a transcript offers to delete its audio file too.
From one set of notes you can generate an HTML report, an HTML slide deck, an editable .pptx, or an .eml email draft — same content, different skins.
The sidebar can organize recent notes by folder or by time. Folder groups can be collapsed, the open note is highlighted, and search and template filters remain available in either view.
Live-outline text can be selected and copied. On desktop, drag a Markdown file from the note list, or hold Alt to drag its path. File acceptance depends on the target application; older hosts may support path dragging only.
- Open the LexVoice sidebar.
- Choose a template and an audio input.
- Start recording; check that the level meter reacts.
- Watch the live outline; add in-meeting notes if needed.
- Stop recording and follow transcription and AI organization in Task progress.
- Ask follow-up questions from Ask this note, or retry only the failed stage if processing was interrupted.
- Open Sediment and review people, todos, learning cards, and hotwords.
- If you need to share, generate an HTML report, slides, PPTX, or an email draft.
Default folders (all configurable in settings):
| Content | Path |
|---|---|
| Recordings | LexVoice/录音 |
| Transcribed notes | LexVoice/转写纪要 |
| Meeting materials | LexVoice/会议资料 |
| People | LexVoice/人员 |
| Learning cards | LexVoice/学习卡片 |
| Todo cards | LexVoice/待办卡片 |
| Views | LexVoice/视图 |
| HTML reports | LexVoice/HTML报告 |
| Email drafts | LexVoice/邮件草稿 |
| Glossary | LexVoice/词汇表.md |
Required:
- Obsidian 1.10.0 or later
- A speech-to-text service (cloud API or local)
- A vault folder for recordings and notes
Recommended:
- An LLM service — powers the live outline, summarization, sediment, export, and template tuning
- A virtual audio device on macOS / Linux, or as a Windows fallback — to record system / online-meeting audio
- A real microphone — to mix in your own voice
- A domain glossary — greatly improves recognition of names, products, organizations, and terms
LexVoice can capture the Windows default playback device directly through Electron Loopback. Choose Microphone + Windows system audio for meetings, or Windows system audio only for video and courses. The capture request temporarily obtains a display stream as Electron requires, immediately discards its video track, and records audio only.
Other desktop platforms, and Windows installations where the direct test fails, can use a virtual audio device:
- Windows fallback: VB-Cable
- macOS: BlackHole
- Linux: PulseAudio / PipeWire monitor source
On Windows with VB-Cable, mind the naming:
- Meeting apps, browsers, and system output → CABLE Input
- LexVoice reads CABLE Output (a recording device)
- To also record yourself, the real microphone must be your physical mic — not CABLE Output, BlackHole, VoiceMeeter, or Stereo Mix
Run Test device before a long recording. A Windows Loopback track can be valid while silent, so play a short piece of audio during the test if you also want to verify the level meter.
No ads, no analytics, no telemetry. Settings are stored locally in .obsidian/plugins/lexvoice/data.json. Recordings are saved to the local vault path you choose; LexVoice has no cloud storage and uploads nothing to any LexVoice server.
However, if you use a cloud ASR or LLM provider, the relevant audio, transcript text, and prompt context are sent to that provider you configured. For sensitive content (client data, medical, legal, HR, recruiting, internal strategy), prefer local transcription + a local model, and obtain consent before recording. See PRIVACY.md.
Automatic links and meeting topics retain the existing setting: enabled by default, with a saved disabled setting preserved. New notes reuse organization requests for relationship extraction and may include up to 40 existing topic names. Background analysis of older/edited notes or missing results sends note IDs, titles, dates, summaries, up to ten topic headings and bounded excerpts, plus up to 24 existing topics with their IDs, names and descriptions. Descriptions may contain information derived from earlier meetings. Topic analysis does not additionally send audio, vault paths or full verbatim transcripts. Provider charges may apply. Enabling the feature allows wikilink, property, MOC and Canvas updates. Disabling it in Library stops subsequent processing but does not undo previously written content.
From the Obsidian Community plugins directory:
- Open Settings → Community plugins → Browse.
- Search for LexVoice, then choose Install and Enable.
- Open LexVoice settings and configure your transcription service and audio input.
Manual install:
- Download
main.js,manifest.json, andstyles.cssfrom the same official GitHub Release. - Copy them into
<your vault>/.obsidian/plugins/lexvoice/. - Reload Obsidian and enable LexVoice under Community plugins.
When updating, keep your existing data.json and vault files. Replace only the three installation files above.
Beginning with the 2.2.0 release line, new LexVoice material is distributed under the LexVoice Proprietary Software License. Development source code is no longer publicly released. Official JavaScript runtime files remain inspectable; that does not make this an open-source release or grant a right to redistribute the runtime or source code.
Personal and internal business use is permitted. Your recordings, notes, and exported reports may still be edited, shared, and commercially used, subject to rights in their contents. Unauthorized rebranding, repackaging, external redistribution (free or paid), resale, and white-label software offerings are prohibited. Local adjustments for your own permitted use are allowed; distributing a modified plugin is not.
Previously MIT-licensed releases, including 2.1.2, and previously MIT-licensed portions reused in later releases retain their original rights. See the preserved MIT notice. Third-party components retain their own licenses; see Third-Party Notices. Nothing here restricts mandatory legal rights or independent implementations of general ideas.
Closed-source distribution through the Obsidian community directory is subject to Obsidian's case-by-case review. This license change does not itself establish approval for the new distribution model.
The HTML slide-deck feature was inspired by alchaincyf/huashu-design; its HTML-first slide workflow and design principles influenced this work. Per the upstream license: Derived from alchaincyf/huashu-design.