Repository navigation
Releases
0.2.2
Changed
- The highlight colour is now Purple by default for new installs. If you already use Edge Voice, your colour stays as it is (Amber unless you picked another); change it under Settings → Highlight colour.
- New demo videos, GIFs and screenshots in the README and docs: dark theme, Purple, and dictation started with Left ⌘ and stopped by itself or with a single tap.
Fixed
- Stop after silence no longer stops dictation while you're still talking. It used to count time since the text last changed, so on a slower computer, where the live text can lag for several seconds, it could cut you off mid-sentence. It now waits until the microphone is also quiet, and the green Done signal does too.
- The green Done signal no longer flashes on and off on a slower computer while the last words are still being recognized: it waits until the recognizer has caught up with everything you said.
SHA-256 of the plugin files:
dfb054f9d7307e375f3bfa79d4d4bfe5ede34731bc41af016760466641a3c070 main.js
50e63ed9563091b656399c2c30859ce9a47caf16ca3d77686774f3111fbb9a91 manifest.json
48d138314c2f439554da6c81f2d8328e51a4751699474b969cc9729bd78087ca styles.css
0.2.1
Added
- Dictation shows when your text is done: once everything you said has been recognized and nothing changes for a second, the microphone gets a pulsing green ring and the status bar turns green with Done · stops in 4 s, counting down to the automatic stop (Done · tap to stop when that's off). Speaking again goes straight back to Dictating. With Reduce motion, the ring and bars stay still.
Changed
- Dictation listens straight away: the microphone opens as soon as you press it, and what you say while the recognizer is still loading is kept and written in once it's ready (the engine still confirms it can dictate first, so an older engine never asks for the microphone).
- When dictation stops, the cursor waits one space after the dictated words, ready for you to type or dictate on.
- Stop after silence now defaults to 6 seconds (was 10). If you had 0.2.0's default of 10, it becomes 6; any other value you chose is kept.
Fixed
- Settings: Get another voice no longer shows a tall empty gap above its controls, and no row's text is squeezed beside its buttons at narrow widths (0.2.0's layout fix stretched some rows).
SHA-256 of the plugin files:
f237c86faa961b746b8076a653463ddf501547e8b08c1a491fd65873dfa5c793 main.js
70fc98a04d4a842ee9f8168505e211151d961da1ce506623833198f73cc99c20 manifest.json
48d138314c2f439554da6c81f2d8328e51a4751699474b969cc9729bd78087ca styles.css
0.2.0
Added
- English dictation (speech-to-text) into Markdown notes, on your own computer, with Moonshine's Tiny, Small or Medium streaming models. Click the microphone in the player, or double-tap Left ⌘ (Mac) / Left Ctrl (Windows, Linux) to start and tap once to stop. Live text appears as you speak and is revised as the recognizer catches up; it stops rather than overwrite anything you type. Optional start/stop sounds, lowercase output and an inactivity timeout. Reading aloud and dictation run one at a time.
- Dictation is on for new installs. If you're updating, it stays off until you turn it on under Settings → Edge Voice → English dictation, so nothing new appears in your player or shortcuts; your reading settings and voices are kept.
- Dictation needs speech engine 0.4.0 and one English model, downloaded with a command the settings show (
edge-voice-engine models get <tiny|small|medium>). Model files are pinned by revision, size and SHA-256 and checked again before each use. Edge Voice still installs and downloads nothing itself; engine 0.3.0 keeps reading aloud. - Edge Voice tells you once when a newer recommended engine is available, with manual update steps. The check is local; nothing is looked up online.
- Matching shortcut pickers for reading and dictation, with conflict checks. One short tap of a mode's key now stops it (double-tap still starts); the player's pause/resume is unchanged.
Fixed
- Settings rows keep their descriptions readable: when buttons don't fit beside the text (for example Storage → Downloaded voices), they move onto their own line instead of squeezing the text into a narrow column.
- Get another voice keeps its instructions and download command full width when a voice is selected.
SHA-256 of the plugin files:
65f423cfcef6dce67bbf478f5ea51d4c0b26787b89f19e2daacd6e488c65db8c main.js
b614b4785f3a865639f56ba778c41d008c36c78b438b695b845504489063e33e manifest.json
cab00d3bf7fe5f03b29a2db53fcb444abd0d637b55f3c7d4c84e91a4452508f6 styles.css
0.2.0-beta.3
Edge Voice 0.2.0-beta.3
Opt-in English dictation beta for Obsidian BRAT. Stable Community users remain on 0.1.5.
Install with BRAT
In Obsidian, install BRAT and use Add a beta plugin with frozen version based on a release tag. Enter tattorba87/obsidian-edge-voice and select 0.2.0-beta.3. Try it in a separate test vault: it uses the same edge-voice plugin ID and replaces the plugin files in that vault.
Setup and preferences
The speech-engine setup window and README list optional TTS voice and STT model commands for each platform. Medium prioritizes word accuracy; Small and Tiny prioritize speed and lower resource use, with more recognition errors expected. Dictation is enabled for new settings, while an existing off choice is preserved. The disabled notice opens English dictation settings directly.
Set up the local engine
BRAT updates the plugin only. Dictation requires speech engine 0.4.0 and an English Tiny, Small or Medium streaming model. Follow setup and update instructions, then Settings → Edge Voice → English dictation → Get the dictation model. Run the shown command yourself and enable dictation. Existing TTS voices/settings are retained; dictation is enabled by default for new settings; a saved off choice is retained.
Click the microphone to start/finish, or double-tap Left Command (Left Ctrl elsewhere) to start and tap once to stop. TTS and STT run one at a time.
Beta limits and checks
English Markdown editing only. MLX accelerates TTS only. Automated checks passed locally: unit/Python/Obsidian tests, production build and fresh packaged-engine offline TTS → STT → TTS. Additional workflow results are available in Actions. The real-engine demo uses a cleared recording routed as input; physical microphone permission/device behavior and sustained speech on every platform still need tester feedback. Partial text can change; stop and wait for final text before editing or switching notes.
For rollback, select stable 0.1.5 explicitly in BRAT or reinstall stable, then remove beta tracking. Do not assume an automatic downgrade.
Dictation guide · STT demo with audio · License and model notes
0.2.0-beta.2
Edge Voice 0.2.0-beta.2
Opt-in English dictation beta for Obsidian BRAT. Stable Community users remain on 0.1.5.
Install with BRAT
In Obsidian, install BRAT and use Add a beta plugin with frozen version based on a release tag. Enter tattorba87/obsidian-edge-voice and select 0.2.0-beta.2. Try it in a separate test vault: it uses the same edge-voice plugin ID and replaces the plugin files in that vault.
Setup and preferences
The speech-engine setup window and README list optional TTS voice and STT model commands for each platform. Medium prioritizes word accuracy; Small and Tiny prioritize speed and lower resource use, with more recognition errors expected. Dictation is enabled for new settings, while an existing off choice is preserved. The disabled notice opens English dictation settings directly.
Set up the local engine
BRAT updates the plugin only. Dictation requires speech engine 0.4.0 and an English Tiny, Small or Medium streaming model. Follow setup and update instructions, then Settings → Edge Voice → English dictation → Get the dictation model. Run the shown command yourself and enable dictation. Existing TTS voices/settings are retained; dictation is enabled by default for new settings; a saved off choice is retained.
Click the microphone to start/finish, or double-tap Left Command (Left Ctrl elsewhere) to start and tap once to stop. TTS and STT run one at a time.
Beta limits and checks
English Markdown editing only. MLX accelerates TTS only. Automated checks passed locally: unit/Python/Obsidian tests, production build and fresh packaged-engine offline TTS → STT → TTS. Additional workflow results are available in Actions. The real-engine demo uses a cleared recording routed as input; physical microphone permission/device behavior and sustained speech on every platform still need tester feedback. Partial text can change; stop and wait for final text before editing or switching notes.
For rollback, select stable 0.1.5 explicitly in BRAT or reinstall stable, then remove beta tracking. Do not assume an automatic downgrade.
Dictation guide · STT demo with audio · License and model notes
0.2.0-beta.1
Edge Voice 0.2.0-beta.1
Opt-in English dictation beta for Obsidian BRAT. Stable Community users remain on 0.1.5.
Install with BRAT
In Obsidian, install BRAT and use Add a beta plugin with frozen version based on a release tag. Enter tattorba87/obsidian-edge-voice and select 0.2.0-beta.1. Try it in a separate test vault: it uses the same edge-voice plugin ID and replaces the plugin files in that vault.
Set up the local engine
BRAT updates the plugin only. Dictation requires speech engine 0.4.0 and an English Tiny, Small or Medium streaming model. Follow setup and update instructions, then Settings → Edge Voice → English dictation → Get the dictation model. Run the shown command yourself and enable dictation. Existing TTS voices/settings are retained; dictation starts disabled.
Click the microphone to start/finish, or double-tap Left Command (Left Ctrl elsewhere) to start and tap once to stop. TTS and STT run one at a time.
Beta limits and checks
English Markdown editing only. MLX accelerates TTS only. Automated checks passed locally: 234 unit tests, 83 Python tests, 52 Obsidian tests, production build and fresh packaged-engine offline TTS → STT → TTS. Additional workflow results are available in Actions. The real-engine demo uses a cleared recording routed as input; physical microphone permission/device behavior and sustained speech on every platform still need tester feedback. Partial text can change; stop and wait for final text before editing or switching notes.
For rollback, select stable 0.1.5 explicitly in BRAT or reinstall stable, then remove beta tracking. Do not assume an automatic downgrade.
Dictation guide · STT demo with audio · License and model notes
Speech engine 0.4.0
The Edge Voice speech engine 0.4.0. Edge Voice shows these steps when it can't find the engine; they're here too. First install uv (https://docs.astral.sh/uv/getting-started/installation/), then run the lines for your computer, in order. Every package is checked against the hash in the requirements file the second line names; read that file first if you like.
Mac (Apple silicon):
uv venv --clear --managed-python ~/.edge-voice/engine --python 3.12
uv pip install --python ~/.edge-voice/engine --require-hashes --no-build -r https://github.com/tattorba87/obsidian-edge-voice/releases/download/engine-0.4.0/requirements-darwin-arm64.txt
~/.edge-voice/engine/bin/edge-voice-engine voices get kokoro_af_bella
Mac (Apple silicon, with the faster engine; macOS 14 or later):
uv venv --clear --managed-python ~/.edge-voice/engine --python 3.12
uv pip install --python ~/.edge-voice/engine --require-hashes --no-build -r https://github.com/tattorba87/obsidian-edge-voice/releases/download/engine-0.4.0/requirements-darwin-arm64-mlx.txt
~/.edge-voice/engine/bin/edge-voice-engine voices get kokoro_af_bella --faster
Mac (Intel):
uv venv --clear --managed-python ~/.edge-voice/engine --python 3.12
uv pip install --python ~/.edge-voice/engine --require-hashes --no-build -r https://github.com/tattorba87/obsidian-edge-voice/releases/download/engine-0.4.0/requirements-darwin-x64.txt
~/.edge-voice/engine/bin/edge-voice-engine voices get kokoro_af_bella
Linux (x64):
uv venv --clear --managed-python ~/.edge-voice/engine --python 3.12
uv pip install --python ~/.edge-voice/engine --require-hashes --no-build -r https://github.com/tattorba87/obsidian-edge-voice/releases/download/engine-0.4.0/requirements-linux-x64.txt
~/.edge-voice/engine/bin/edge-voice-engine voices get kokoro_af_bella
Linux (ARM):
uv venv --clear --managed-python ~/.edge-voice/engine --python 3.12
uv pip install --python ~/.edge-voice/engine --require-hashes --no-build -r https://github.com/tattorba87/obsidian-edge-voice/releases/download/engine-0.4.0/requirements-linux-arm64.txt
~/.edge-voice/engine/bin/edge-voice-engine voices get kokoro_af_bella
Windows (PowerShell):
uv venv --clear --managed-python "$HOME\.edge-voice\engine" --python 3.12
uv pip install --python "$HOME\.edge-voice\engine" --require-hashes --no-build -r https://github.com/tattorba87/obsidian-edge-voice/releases/download/engine-0.4.0/requirements-win32-x64.txt
& "$HOME\.edge-voice\engine\Scripts\edge-voice-engine.exe" voices get kokoro_af_bella
Windows on ARM (PowerShell):
uv venv --clear --managed-python "$HOME\.edge-voice\engine" --python cpython-3.12-windows-x86_64-none
uv pip install --python "$HOME\.edge-voice\engine" --require-hashes --no-build -r https://github.com/tattorba87/obsidian-edge-voice/releases/download/engine-0.4.0/requirements-win32-x64.txt
& "$HOME\.edge-voice\engine\Scripts\edge-voice-engine.exe" voices get kokoro_af_bella
The macOS moonshine-voice wheels are built by this repository and contain ONNX Runtime and other open-source libraries; their licences are in THIRD-PARTY-NOTICES.txt.
SHA-256 of the files:
8a731b1b4d3fb7bebffacc9f22b3a718f6f3f40da5d7b9c595b4f080cb67465a docopt-0.6.2-py2.py3-none-any.whl
5ee828d30f9e9ae68b391fedbfdaec4ca5e3a8e1ff692f72dd07a82b8db68e87 edge_voice_engine-0.4.0-py3-none-any.whl
5328a49f62ff4670dada9a0bb7f950994a3ba0b2d47898a7049f540d79afbafd moonshine_voice-0.1.5-py3-none-macosx_13_0_arm64.whl
a3a5d827cbb31339808ba0cba0f0082bd26550f0a6a2d5577fc8bf65f382ad23 moonshine_voice-0.1.5-py3-none-macosx_13_0_x86_64.whl
59e2c23c0100ce724d59a4bf28df3ac6da9aa748477e61769d251a7e71e27efe requirements-darwin-arm64-mlx.txt
e72f54f6bf40832545de5c5e6edbfb305fb5fd5b1fe1127f08a3227e1f35c788 requirements-darwin-arm64.txt
631039c56da814620d36c90a501930b99a837f301bc95b84c0da853eebf6a986 requirements-darwin-x64.txt
4718c6211a3c34e63938d0fefba6c2df8840f410bdbca332f2a754cd10f86fac requirements-linux-arm64.txt
4718c6211a3c34e63938d0fefba6c2df8840f410bdbca332f2a754cd10f86fac requirements-linux-x64.txt
211a4845485e8b014a3344d5ca16e8a14ffa8e8d25c79e158ebe55c0d0c9b14b requirements-win32-x64.txt
604844f22b6b700488d281c1bc5319ca2d14ebd955d945e8bddafb8260bd57f8 THIRD-PARTY-NOTICES.txt
Optional English dictation models
The same engine supports TTS reading and English STT dictation. Download a TTS voice for reading aloud, an STT model for dictation, or both. The model command is separate from voices get; skip the voice command if you only want dictation. The plugin and BRAT never install Python or models for you.
After installing this engine, run one of these commands yourself:
macOS / Linux:
~/.edge-voice/engine/bin/edge-voice-engine models get mediumWindows PowerShell:
& "$HOME\.edge-voice\engine\Scripts\edge-voice-engine.exe" models get mediumMedium (269 MB) prioritizes word accuracy. Replace medium with small (142 MB) or tiny (45 MB) to prioritize speed and lower resource use, with more recognition errors expected, especially with Tiny. All three are English Moonshine streaming models under MIT. Files download from download.moonshine.ai at a pinned revision and are checked by size/SHA-256.
In Edge Voice's English dictation settings, select the model you downloaded and leave Custom model folder empty. Check Enable dictation is on; a saved off preference is retained. Click the microphone or use its shortcut to start recording. MLX/faster voice packages accelerate TTS only and are not required for STT.
0.1.5
Added
- Settings → Advanced → Free voice memory when idle: after 5, 10 or 30 minutes or an hour stopped or paused, the voice engine is unloaded to give its memory back (about 800 MB with a Kokoro voice). The status bar shows "Unloading voice", then "Reloading voice" on the next play, resume or skip (about 2–3 seconds). Off by default.
SHA-256 of the plugin files:
0ad216bd2cd4c59fd02da1dc9916d9284c73e29b205774dadb8b245b202209f6 main.js
dd4aa00bd1cd157365262c7d74f899e4d848b22172fc6d4ed7987558efda68f0 manifest.json
30668c4f066ba33ce48aa06531d9eb7f09a3ebd7886b3eb260a455092fa3214a styles.css
0.1.4
Fixed
- In Reading view and other rendered text, a sentence whose first word also appears a few words earlier in the paragraph ("…your own computer. Your notes…") was highlighted from the earlier word, and its word highlight ran ahead.
- Hyphenated words are said properly: "text-to-speech", "step-by-step" and "chain-of-thought" were run together.
- "us" was said as the letters "U S".
- "read" is said "red" after "has", "have", "had", "been", "was", "were", "be", "well" and "widely" ("it was read aloud"); elsewhere it stays "reed".
Changed
- New demo: why reading along helps, and a research paper read aloud.
SHA-256 of the plugin files:
e94b034a794841370b7480d1a985771c06cbe67da1afd042050f409c400c7f40 main.js
9de8645ca89b8d883966215dee6c23a453ad01467a15e5a2fdb0bb456a498dbb manifest.json
30668c4f066ba33ce48aa06531d9eb7f09a3ebd7886b3eb260a455092fa3214a styles.css
0.1.3
Changed
- Edge Voice's settings show up in Obsidian's settings search (Obsidian's setting definitions, 1.13+). Storage and Advanced are now sections with headings instead of folded away, so search can reach every row.
- Renamed the engine connection's internal
requestmethod, which Obsidian's scan counted as network requests. The plugin makes no network requests.
Fixed
- Settings: the custom double-tap key box stayed visible while the shortcut was off.
SHA-256 of the plugin files:
aa2f24679b1e4163fe76a6415770b42733a280c34f3ef9198d4c543f4772f440 main.js
858716e315e0774ad3f04b4506cabeb16202b8236c0621ce3d33428fb159fa13 manifest.json
30668c4f066ba33ce48aa06531d9eb7f09a3ebd7886b3eb260a455092fa3214a styles.css