Local Speech Recognition
Alfred Clark185 downloadsUse the locally deployed sherpa-onnx service to realize offline speech recognition in your notes.
- Overview
- Scorecard
- Updates9
English | 简体中文
Offline speech recognition in Obsidian, powered by a locally deployed sherpa-onnx service. Press a hotkey, speak, and the transcription is inserted at your cursor. No cloud, no API keys.
Features
- Offline transcription via
sherpa-onnx-offline-websocket-serverover a local WebSocket (defaultws://127.0.0.1:6006). - Hotkey-triggered recording with toggle and push-to-talk modes. The plugin ships with no preset hotkey; assign one yourself in Settings → Hotkeys.
- Microphone selection with device enumeration and one-click refresh.
- 16 kHz 16-bit mono pipeline:
AudioWorkletcapture with linear downsampling. - Smart delivery: inserts at the cursor when a Markdown editor is focused (source mode), otherwise copies to the clipboard with a Notice.
- Lexicon (designed for Chinese): add words from the editor context menu; after a transcription, matching words are underlined and can be clicked to replace with a same-pinyin alternative. Optional fuzzy pinyin matching tolerates retroflex/nasal confusions.
- Sidebar view (ribbon icon): a Lexicon tab for managing entries and a Service tab for controlling the service.
- Service management: manual start/stop/restart, auto-start on plugin load, and a connection test.
- Selectable recognition models: SenseVoice, FunASR, Paraformer and Qwen3-ASR families with int8/fp16/fp32 variants, each started with its own parameters (see Supported models).
- Trilingual UI: English / 简体中文 / 繁體中文, following the Obsidian interface language or set manually.
- Gamepad control (just for fun): drive recording, cursor, candidate picking and more with a standard-mapping controller (see Gamepad control).
Requirements
- Obsidian 1.13.1 or later (uses the declarative settings API).
- Desktop only. The plugin spawns a local subprocess and captures microphone audio, neither of which is available on mobile.
- A
sherpa-onnx-offline-websocket-serverbinary (see Choosing the server binary) plus a model folder holding one of the supported models (see Supported models).
Downloads
- sherpa-onnx releases (server binaries): https://github.com/k2-fsa/sherpa-onnx/releases
- Pre-trained ASR models: https://github.com/k2-fsa/sherpa-onnx/releases/tag/asr-models
Choosing the server binary
Each sherpa-onnx release ships pre-built archives for desktop platforms. Pick the one matching your OS and CPU, extract it, then point Settings → Service settings → Server binary path at the executable inside its bin/ folder:
- macOS / Linux:
<extracted-folder>/bin/sherpa-onnx-offline-websocket-server - Windows:
<extracted-folder>/bin/sherpa-onnx-offline-websocket-server.exe
| Platform | Archive to download |
|---|---|
| macOS (Apple silicon) | sherpa-onnx-v<version>-osx-arm64-static.tar.bz2 |
| macOS (Intel) | sherpa-onnx-v<version>-osx-x64-static.tar.bz2 |
| macOS (both) | sherpa-onnx-v<version>-osx-universal2-static.tar.bz2 |
| Linux (x86_64) | sherpa-onnx-v<version>-linux-x64-static.tar.bz2 |
| Linux (ARM64) | sherpa-onnx-v<version>-linux-aarch64-static.tar.bz2 |
| Windows (x64) | sherpa-onnx-v<version>-win-x64-static-MD-Release.tar.bz2 |
| Windows (ARM64) | sherpa-onnx-v<version>-win-arm64-static-MD-Release.tar.bz2 |
| Windows (32-bit) | sherpa-onnx-v<version>-win-x86-static-MD-Release.tar.bz2 |
Notes:
- On Linux, when sherpa-onnx was installed through a package manager (pacman, yay, …) rather than a release archive, locate the installed binary with
whereis sherpa-onnx-offline-websocket-serverand use that path. - Prefer the
-staticarchives: their executable is self-contained. The-sharedvariants additionally need the archive'slib/folder on the library search path (PATHon Windows,LD_LIBRARY_PATHon Linux,DYLD_LIBRARY_PATHon macOS). - Append
-no-ttswhere available (e.g....-linux-x64-static-no-tts.tar.bz2): the plugin only transcribes speech, and those archives are much smaller. - macOS may refuse to run the unsigned binary when a browser set the quarantine flag; clear it with
xattr -dr com.apple.quarantine <extracted-folder>.
Supported models
Pick a model under Settings → Service settings → Recognition model, then point Model folder path at the folder you extracted. That folder must contain the files listed below; extra files such as test_wavs/ are ignored. Switching models takes effect the next time the service starts.
| Recognition model | Model folder must contain | Characteristics |
|---|---|---|
| SenseVoice (int8) | model.int8.onnx, tokens.txt |
Default. Mandarin, Cantonese, English, Japanese and Korean; ITN adds punctuation; small (~230 MB) and fast |
| SenseVoice | model.onnx, tokens.txt |
The same model at full precision (~890 MB): slightly more accurate, but larger and slower |
| FunASR (int8) | encoder_adaptor.int8.onnx, llm.int8.onnx, embedding.int8.onnx, Qwen3-0.6B/ |
Uses Qwen3-0.6B as the decoder, the most accurate option here; Mandarin (7 dialect groups, 26 regional accents), English and Japanese, and it also handles sung lyrics. Large (~950 MB) and slow |
| FunASR (fp16) | encoder_adaptor.int8.onnx, llm.fp16.onnx, embedding.int8.onnx, Qwen3-0.6B/ |
FunASR with an fp16 decoder (~1.5 GB); a middle ground between the int8 and fp32 builds |
| FunASR (fp32) | encoder_adaptor.onnx, llm.fp32.onnx, llm.fp32.data, embedding.onnx, Qwen3-0.6B/ |
FunASR at full precision (~3.7 GB), the heaviest option; keep llm.fp32.data next to llm.fp32.onnx |
| Paraformer (int8) | model.int8.onnx, tokens.txt |
Non-autoregressive and the fastest model here (~220 MB); Chinese (some builds add English or Cantonese); the output carries no punctuation |
| Paraformer | model.onnx, tokens.txt |
Paraformer at full precision (~790 MB), the largest of its family |
| Qwen3-ASR 0.6B (int8) | conv_frontend.onnx, encoder.int8.onnx, decoder.int8.onnx, tokenizer/ |
LLM-based, multilingual with punctuation (~1 GB). Fast for its size, occasional errors on tongue-twisters; CPU recommended for int8 (see note below) |
| Qwen3-ASR 1.7B (int8) | conv_frontend.onnx, encoder.int8.onnx, decoder.int8.onnx, tokenizer/ |
Same layout as 0.6B (~2.4 GB). More accurate on normal speech but slower, and may truncate highly repetitive input; CPU recommended for int8 (see note below) |
tokenizer/ is a folder, not a single file. Qwen3-0.6B/ (FunASR) is likewise a tokenizer folder. Archives that match the table, from the ASR models release:
- SenseVoice:
sherpa-onnx-sense-voice-zh-en-ja-ko-yue-2024-07-17(contains bothmodel.onnxandmodel.int8.onnx) orsherpa-onnx-sense-voice-zh-en-ja-ko-yue-int8-2024-07-17(int8 only). - FunASR:
sherpa-onnx-funasr-nano-int8-2025-12-30,sherpa-onnx-funasr-nano-fp16-2025-12-30, andsherpa-onnx-funasr-nano-2025-12-30for fp32. - Paraformer:
sherpa-onnx-paraformer-zh-2024-03-09and other archives of the same family —-zh-small-2024-03-09is the smallest,-trilingual-zh-cantonese-ensupports Cantonese,-zh-int8-2025-10-07targets Sichuanese. Some builds (small, Sichuanese) shipmodel.int8.onnxonly, so pair them with Paraformer (int8). - Qwen3-ASR:
sherpa-onnx-qwen3-asr-0.6B-int8-2026-03-25for 0.6B. There is no official int8 package for 1.7B yet; assemble its folder frommodel_1.7Binzengshuishui/Qwen3-ASR-onnx(ModelScope) using the same file layout as the 0.6B package.
Note: a model of the same architecture with different file names can in theory be forced to run by renaming its files to the names listed above, but this is not guaranteed to work.
Note: Qwen3-ASR runs on CPU out of the box. CUDA works only when the server binary and cuDNN versions match — e.g. a system-packaged binary expecting cuDNN 8 crashes at startup on cuDNN 9 systems, while a matching build runs fine. For int8 models CPU is as fast or faster, so CUDA buys nothing there.
Installation
- In Obsidian, open Settings → Community plugins → Browse, search for "Local Speech Recognition", and install it. After installation, enable the plugin.
- If you prefer a manual install, download
main.js,manifest.json, andstyles.cssfrom the GitHub release, copy them to<vault>/.obsidian/plugins/local-speech-recognition/, reload Obsidian, then enable the plugin in Settings → Community plugins.
Quick start
- Download the server binary and the model from the links above, and unpack the model folder somewhere local.
- Open plugin settings → Service settings: fill in the server binary path and model folder path, check host/port/threads, then click Test connection or Start service (or enable auto-start).
- Open Speech input: pick a microphone. If the list is empty or unlabeled, allow microphone access once, then click Refresh microphones. Choose an input mode.
- Open Obsidian Settings → Hotkeys, find "Toggle speech recognition", and assign a hotkey.
- Focus a note editor, press the hotkey, speak, then press it again. The text is inserted at the cursor. If no editor is focused, the result goes to the clipboard with a Notice instead.
- Optional: build a lexicon of names or jargon. Select a word in an editor and pick Add to current lexicon from the right-click menu (or Add to global/vault lexicon to write directly to the other storage without switching); later transcriptions will underline it and offer one-click corrections. See Lexicon below.
Settings reference
General:
| Setting | Description |
|---|---|
| Plugin language | Interface language: follow system, English, 简体中文, 繁體中文 |
| Collapsible groups | Render settings groups as navigable sub-pages when on |
Service settings:
| Setting | Description |
|---|---|
| Server binary path | Path to the sherpa-onnx-offline-websocket-server executable |
| Model folder path | Folder holding the files required by the selected model (see Supported models) |
| Recognition model | Which supported model the service loads at start; takes effect on the next start |
| Validate paths | Check the server binary and the selected model files before starting |
| Host / Port | Where the service listens (defaults 127.0.0.1 / 6006) |
| CPU threads | Threads used by the service neural network |
| Test connection | Dial the configured WebSocket address to check reachability |
| Auto start server | Start the service automatically when the plugin loads |
| Start/Stop/Restart | Manual lifecycle controls, shown depending on service status |
Speech input:
| Setting | Description |
|---|---|
| Microphone | Input device (empty means the system default); the refresh button re-enumerates devices |
| Input mode | Toggle on click, or push-to-talk (see below) |
| Status notification | Notice, status bar, or none for progress; errors always use notices |
| Auto-stop | Stop and transcribe automatically after the silence duration (off by default) |
| Silence duration | Seconds of silence to wait before auto-stop, shown only when auto-stop is on (1–5 s, default 2 s) |
Lexicon settings:
| Setting | Description |
|---|---|
| Enable lexicon | Turn all lexicon features on or off, including the sidebar tab and editor integration (default on) |
| Fuzzy pinyin matching | Also match words differing only in retroflex initials or front/back nasals (default off; see Lexicon below) |
| Transcription history | Keep recognition highlights across insertions: none, recent N, or all (default none) |
| History count | How many recent insertions to keep, shown only for recent-N mode (1–20, default 3) |
| Export / Import / Clear | Back up to JSON, import from JSON or one-word-per-line text, or permanently delete every entry (hidden while disabled) |
Gamepad settings (all hidden while the gamepad is disabled; see Gamepad control):
| Setting | Description |
|---|---|
| Enable gamepad | Turn gamepad input on or off (default off) |
| Stick dead zone | Ignore stick drift below this value (default 25%, range 0–50%) |
| Scroll speed | Viewport scroll multiplier (default ×1.0, range ×0.5–×2.0) |
| Invert scroll direction | Flip the scroll stick vertically (default off) |
| Character step interval | Step interval for horizontal stick moves, also the full-press trigger baseline (default 60 ms, range 20–200 ms) |
| Line step interval | Step interval for vertical stick moves and line jumps (default 120 ms, range 40–500 ms) |
| Hold repeat delay | Shared initial delay before direction pad, line jump, backspace and triggers repeat (default 400 ms, range 100–1000 ms) |
| Direction pad interval | Repeat step interval while holding a direction pad direction (default 130 ms, range 40–400 ms) |
| Backspace interval | Repeat step interval while holding backspace (default 60 ms, range 20–200 ms) |
| Trigger threshold | Trigger depth below this value counts as released (default 15%, range 5–40%) |
Usage notes
- Toggle mode: press once to start, press again to stop and transcribe.
- Push-to-talk mode (transitional): press to start; releasing the hotkey or pressing again stops and transcribes. Obsidian hotkeys only deliver key-down events, so exact hold-to-talk semantics are approximated for now.
- Recording is capped at 280 seconds (the server rejects utterances over 300 seconds) and then transcribed automatically.
- While a transcription is in flight, new triggers are rejected with a Notice.
- Lexicon: select text in an editor and choose Add to current lexicon from the right-click menu. After that, every transcription is scanned for lexicon words; matches get a dotted underline, and clicking one opens a menu of same-pinyin replacements.
- Manage entries in the sidebar (ribbon icon) → Lexicon: search, filter by enabled state, add/edit/delete, and batch enable/disable/delete.
Lexicon
The lexicon is designed for Chinese. Each entry pairs a word with its pinyin, and matching compares the pinyin of transcribed text against those entries. That means a word can still be flagged when the speech model returns a homophone — e.g. the lexicon knows 知识, so a transcription of 资石 (identical pinyin zhi shi) is underlined and can be corrected with one click. Because matching is pinyin-based, the lexicon has no effect on non-Chinese text: Latin-script segments are ignored, and words without a pinyin are skipped.
Enable Fuzzy pinyin matching to also treat words as equivalent when they differ only in retroflex initials (zh/z, ch/c, sh/s) or front/back nasals (an/ang, en/eng, in/ing, ian/iang, uan/uang). This helps with accents and model confusions, but it trades precision for recall — with it on, 张 and 脏 become mutually interchangeable — so it is off by default.
Notes:
- Pinyin is generated automatically when you add a word; you can edit it, or regenerate it from the word, in the entry dialog.
- An entry can be disabled instead of deleted; disabled entries are kept in the list but excluded from matching.
- Assigning a weight ranks entries that share a pinyin, so higher-weight words appear first among the replacements.
- Adding a word that already exists (same word and pinyin) is rejected with a Notice; homophones with different characters are allowed.
Gamepad control
Just for fun: gamepad support is a playful extra, not a productivity feature. It assumes a standard-mapping controller (Xbox-style) on desktop and a focused Markdown editor in source mode; most actions stay silent when the editor is not focused. Enable it under Settings → Gamepad settings → Enable gamepad — a fixed key map for reference is shown right below the toggle.
| Button | Candidate popup open | Editing text (focused editor) |
|---|---|---|
| Y | Close the popup, then toggle recording | Toggle recording (works unfocused; result goes to clipboard) |
| A | Confirm the highlighted candidate, stay put | Insert a newline |
| B | Dismiss the popup (single press, no repeat) | Backspace, hold to repeat-delete |
| X | Close the popup and undo | Undo the last edit (single fire, no repeat) |
| LB / RB | Close the popup, jump prev / next logical line (hold to repeat, column preserved) | Same as left |
| LT / RT | Start or continue extending the selection backward / forward | Same as left; depth controls speed, release keeps the selection, reversing shrinks it |
| D-pad ← / → | Switch to the prev / next candidate span | Move by character (hold to repeat) |
| D-pad ↑ / ↓ | Switch candidates inside the popup | Jump visual lines incl. wrapped lines (hold to repeat) |
| Left stick | Scroll the viewport vertically (no focus needed) | Same as left; speed and direction follow the scroll settings |
| Right stick | X: move by character; Y: move by visual line | Same as left |
Notes:
- Triggers are analog: pressing deeper repeats faster, from twice the character interval down to the full-press baseline; below the trigger threshold counts as released.
- Holding any repeating input first fires once, then repeats after the shared hold delay.
- Editing actions (backspace, newline, undo, confirmed replacement) exit candidate navigation; pure cursor moves only close the popup.
Troubleshooting
| Symptom | What to do |
|---|---|
| Microphone list empty or unlabeled | Allow microphone access once, then click Refresh microphones |
| "Service is not running" Notice | Start the service from settings, or enable auto-start |
| Connection test fails | Check binary/model paths, host/port, and whether the port is in use |
| Microphone denied / not found | Check OS microphone permission and that the selected device is plugged in |
| Recognition fails | Check the Notice detail; for exit codes, verify the model folder layout |
| Service fails to start with missing paths | The Notice names the missing files; match the model folder to the selected model |
| An unexpected word gets underlined | Fuzzy pinyin matching is on; turn it off under Lexicon settings |
Privacy
All recognition runs locally. Microphone audio is sent only to the WebSocket address you configure (default localhost) and never to any third party.
Development
bun install
bun run dev # watch + rebuild into dist
bun run build # typecheck + production build
bun run lint
bun run format
License
GPL-3.0-only. See LICENSE.
HealthExcellent
ReviewSatisfactory
About
Transcribe speech offline via a local sherpa-onnx WebSocket server and insert transcriptions at the cursor in the Markdown editor. Use hotkey-triggered recording (toggle or push-to-talk), select a microphone, manage the local service, and copy output to the clipboard when the editor isn't focused. Run on desktop with a local sherpa-onnx server and compatible int8 model files.Details
Current version
0.6.0
Last updated
4 weeks ago
Created
2 months ago
Updates
9 releases
Downloads
185
Compatible with
Obsidian 1.13.1+
Platforms
Desktop only
License
GPL-3.0