Repository navigation
Releases
Sunoto 0.1.17
New installs are ready to use sooner, and Macs that have had Sunoto since its earliest versions now get the word list.
- Faster first start. Setup no longer waits for the speech model before it opens the app. The permission panel opens right away, and the model downloads in the app, with progress in the panel, the pill and the menu bar. The download is now 851 MB instead of 1.2 GB, it's unpacked while it downloads, and an interrupted download picks up where it left off.
- Word list for older installs. Installs from 0.1.9 or earlier were still on an older speech engine that re-transcribed the whole recording after every dictation and never used the word list. They now move to the current engine automatically. Names and terms on your screen help transcription, long dictations are transcribed while you talk, and the wait after the shortcut no longer grows with length. Nothing to do: it happens on update.
Sunoto still runs on macOS 14 or later.
0.1.11 and later update themselves. If you have 0.1.10 or earlier, install or upgrade with:
curl -fsSL https://raw.githubusercontent.com/shakdwipeea/voice-dictation/master/install-macos.sh | bashValidation:
- cargo test, clippy and the Python suites pass.
- On a Mac, the published 0.1.16, set to the old speech engine, updated itself to this version. Its config moved to the current engine, it came back ready with the engine running from the installed app, and it recovered when that engine was killed.
- The release build's app, overlay and llama.cpp libraries all require macOS 14.0, as in 0.1.16.
Sunoto 0.1.16
Sunoto now has a proper Settings window and a new menu bar panel, so you can set it up and control it without touching a config file.
- Settings. Choose your shortcut and whether you tap or hold it. Pick a microphone and test it. See your permissions, turn Open at Login on or off, edit your vocabulary, and decide how long history is kept and which apps are left out.
- Menu bar panel. Click the Sunoto icon to see your recent dictations and click one to copy it. You can also report a wrong dictation, pause Sunoto, check for updates, and turn automatic updates on or off. The app you're writing in keeps focus while the panel is open.
- Pause. Pause turns the shortcut off until you resume.
- Practice and setup guide. Try dictating in a private window where nothing gets pasted, or walk through setup again at any time.
Fixes:
- Your clipboard is restored if Sunoto shuts down in the middle of a paste.
- A shortcut tap can no longer get stuck waiting after you release the key.
- Practice now ends properly when you leave it.
- The
sunoto-daemoncommands no longer fail when the app takes a moment to read them.
Sunoto 0.1.15
You can now tell us when Sunoto gets a dictation wrong, straight from the menu bar. Sunoto also installs itself the first time you open it, with no Terminal needed.
-
This Was Wrong… Open Recent Dictations in the menu bar, pick a dictation, and choose This Was Wrong…. A panel lists exactly what the report holds:
- your Sunoto and macOS versions and Mac model;
- how long the recording was and how fast it was transcribed;
- your microphone type;
- how many words the word list took from the screen and the project. It sends counts only, never the words.
The text of the dictation goes only if you tick Include the text. Click Send and you get a short code such as
SUN-7F3Ato quote if you write to us. Nothing is sent unless you click Send. If you're offline, the report goes once you're back online. -
Opens like any other app.
- Opened from Downloads, Sunoto offers to move itself to Applications.
- It adds itself to Login Items without a system prompt.
- It downloads its speech model in the app, with progress shown in the menu bar and the pill.
0.1.11 and later update themselves. If you are on 0.1.12 and Sunoto shows "ASR backend lost", restart your Mac first; it then updates on its own. If you have 0.1.10 or earlier, install or upgrade with:
curl -fsSL https://raw.githubusercontent.com/shakdwipeea/voice-dictation/master/install-macos.sh | bashValidation:
- cargo test, clippy and the Python suites pass.
- On a Mac, the published 0.1.14 updated itself to this version. It came back ready, its speech engine ran from the installed app, and it recovered when that engine was killed.
- Reports were sent from the real menu and checked in Sentry:
- with and without the text;
- offline, where the report waited and went within a minute of the network returning;
- from two real dictations, whose text matched what was said.
Sunoto 0.1.14
Sunoto now has an icon and lives in your menu bar. You can see that it's ready, find what you dictated, and manage updates without opening a terminal.
- A menu bar icon. A small waveform sits in the menu bar. Click it to see:
- whether Sunoto is ready (green dot) or still loading, and which key to tap to dictate;
- Recent Dictations: your last 10 dictations and how long ago you made them. Click one to copy its text. Clear History… deletes your dictation history after asking you first. Password fields and password managers never appear here.
- Check for Updates…: tells you that you're up to date, or offers the new version with Install Now or Later.
- Install Updates Automatically: turns automatic updates off or on with one click. It's on unless you turn it off.
- the version you're running.
- An app icon. Sunoto now has its own icon in Finder, in Login Items, and in Privacy & Security, where you grant its permissions, so it's easy to find.
0.1.11 and later update themselves. If you are on 0.1.12 and Sunoto shows "ASR backend lost", restart your Mac first; it then updates on its own. If you have 0.1.10 or earlier, install or upgrade with:
curl -fsSL https://raw.githubusercontent.com/shakdwipeea/voice-dictation/master/install-macos.sh | bashValidation:
- cargo test (265), clippy and the Python suites pass, and the menu bar app compiles without warnings.
- On a Mac, a 0.1.13 release build updated itself to this version and came up with the menu bar and icon.
- Clicked in the real menu, across two more test versions:
- Check for Updates → Install Now installed the next version at once.
- With Install Updates Automatically unticked, nothing installed.
- Ticking it again installed the waiting version within 25 seconds.
- Recent Dictations showed new dictations, copied their text, and Clear History emptied it.
Sunoto 0.1.13
This release fixes "ASR backend lost; restarting" after the update to 0.1.12.
- Fixed: dictation stopped working after the update to 0.1.12. When Sunoto updated itself, the new version could start from the update's temporary download instead of from the installed app. The download was cleaned up right after installing, so the next time the speech engine restarted it could not start, and every dictation ended with "ASR backend lost; restarting". Sunoto now always runs from its own installed copy, however it was started.
- Updates reach a broken app. If the speech engine cannot start, Sunoto now still installs a waiting update after 10 minutes without dictation, so a future fix can arrive on its own.
If Sunoto shows "ASR backend lost" right now, restart your Mac. That fixes it, and Sunoto then updates itself to 0.1.13. A Mac stuck this way cannot install 0.1.13 by itself, because 0.1.12 waits for a working speech engine before it installs an update.
0.1.11 and later update themselves. If you have 0.1.10 or earlier, install or upgrade with:
curl -fsSL https://raw.githubusercontent.com/shakdwipeea/voice-dictation/master/install-macos.sh | bashValidation:
- The 0.1.12 loop was reproduced by starting the app with a runtime folder that was then deleted. The fixed build ignores that folder and its speech engine restarts cleanly.
- The published 0.1.12 release, pointed at a local update feed, updated itself to a 0.1.13 build of this fix in 18 s. The new app ran its speech engine from
~/Applications/Sunoto.app, the download was deleted, and the engine came back after being stopped on purpose. - cargo test (258), clippy, fmt and the Python suites pass.
Sunoto 0.1.12
Sunoto now spells the names you use: your projects, tools, apps, and the names on your screen. It moves to a better English speech model, and on MacBooks the pill grows out of the notch.
- Names spelled right, out of the box. When you start a dictation, Sunoto gathers the unusual words around you: from the window you are typing in, the project you are working on, your project folders and your apps. The speech model leans toward those spellings while it decodes. A near miss ("Tsunoto", "mini sign") is then corrected to the word on the list. There is nothing to set up and nothing to teach. The words are read on your Mac and used for that one dictation only, and they are never saved or logged. In our tests Sunoto got 74 of 93 project and tech names right, up from 11 in 0.1.11, and 395 of 536 names across six kinds of users, up from 207. Speed and memory are unchanged.
- More names if you want them. In
~/Library/Application Support/sunoto/config.json:"word_list_contacts": trueadds the names in Contacts. macOS asks once."word_list_documents": trueadds your Documents and Desktop folder names and recently opened documents. macOS asks once."word_list": ["Kubernetes", "Priya"]adds your own words."word_list_enabled": falseturns the feature off.
- A better English speech model. Sunoto now runs Parakeet TDT 0.6B v2 in place of the multilingual v3. It made fewer mistakes on every English test, at the same speed and memory. On the LibriSpeech rare-word benchmark, 8.6% of rare words came out wrong with no list, down from 10.2%. Two things change:
- English only. If you dictate in another language, set
"asr_model": "mlx-community/parakeet-tdt-0.6b-v3"in the config to keep the old model. Sunoto then downloads it again, 2.5 GB. - Numbers come out as digits: 25 rather than twenty-five.
- English only. If you dictate in another language, set
- The update stays quick. It installs in seconds and keeps the old model for now. Sunoto then downloads the new one in the background (1.24 GB, once) and switches to it after 10 minutes without dictation. Then it deletes the old model it had downloaded, which frees 1.25 GB. It never touches
~/.cache/huggingface, because other apps may use the models there. If you installed before 0.1.11, the old model may still be there ashub/models--mlx-community--parakeet-tdt-0.6b-v3(2.5 GB). Delete it if nothing else uses it. - The pill grows out of the notch on Macs with a camera notch. It is a short island under the camera with the dot, your voice and the timer, and a check mark in the middle after a paste. Screens without a notch keep the floating pill, and
"overlay_style": "pill"keeps it everywhere. - Dictation history. Sunoto keeps a private history of your dictations on this Mac, so you can find what you said. It stores text and timings only, never audio, for 90 days. Password fields and password managers are left out.
"history_enabled": falseturns it off, andsunoto history clear --yesdeletes it. - Fixes. After a paste the check mark lands in the middle of the pill. The pill no longer jitters while you dictate, and messages fit their text. Installs first set up with 0.1.0–0.1.6 now turn the microphone off when a dictation ends, not two minutes later.
0.1.11 updates itself to this version. If you have 0.1.10 or earlier, install or upgrade with:
curl -fsSL https://raw.githubusercontent.com/shakdwipeea/voice-dictation/master/install-macos.sh | bashValidation:
- cargo test (256), clippy -D warnings and the Python suites pass.
- Names were benchmarked with synthetic voices for six kinds of users, with background talk, on the LibriSpeech rare-word benchmark, and on real dictations.
- Qwen3-ASR was also tried and not adopted: it was 2–3× slower and wrote text into silence.
- On a Mac set up like a 0.1.11 install (the v3 copy, no v2), the new
setup --updatefinished in 7 s. The app then downloaded and verified v2 in the background, switched after the idle wait, came back ready in 2.3 s, and deleted v3.
Speech model: Parakeet TDT 0.6B v2 (bf16)
Speech model for Sunoto: a bfloat16 copy of Parakeet TDT 0.6B v2 (English). This is not an app release; install Sunoto from the latest release.
Sunoto's setup downloads these files instead of the 2.5 GB float32 original. The speech engine (parakeet-mlx) converts every weight to bfloat16 when it loads the model, so this is the same model it already runs: all 697 tensors are bit-identical to what it loads from the original, at half the download (1.24 GB).
- Model: Parakeet TDT 0.6B v2 by NVIDIA, https://huggingface.co/nvidia/parakeet-tdt-0.6b-v2, licensed under CC-BY-4.0 (https://creativecommons.org/licenses/by/4.0/).
- MLX conversion: mlx-community, https://huggingface.co/mlx-community/parakeet-tdt-0.6b-v2.
- Change made here: every float32 tensor in
model.safetensorscast to bfloat16 with MLX (round to nearest even), withtools/make_model_copy.py, which also checks the result.config.jsonis unchanged.
Checksums are in SHA256SUMS; setup verifies both files before using them.
Sunoto 0.1.11
Sunoto now updates itself. This is the last version you need to install by hand.
- Automatic updates. Once a day Sunoto checks for a new release. It downloads it and checks its signature against a key built into the app, so only genuine releases install. It installs the update after 10 minutes without dictation, and restarts in a few seconds with your settings and permissions kept. If the new version does not start, the previous one comes back, and that version is not tried again.
sunoto updateinstalls the newest release right away;sunoto update --checkonly looks.- Your choice. Set
"auto_update"in~/Library/Application Support/sunoto/config.jsonto"notify"to only log that an update is out, or"off"to stop checking. The default is"install". Progress is written to~/Library/Logs/sunoto/update.log.
Install or upgrade. Run this once more if you have 0.1.10 or earlier; later versions arrive on their own:
curl -fsSL https://raw.githubusercontent.com/shakdwipeea/voice-dictation/master/install-macos.sh | bashValidation: cargo test (214), clippy, and fmt pass. On a Mac with release builds and a local test feed, an install updated itself from 0.1.10 to 0.1.11 and came back ready with the shortcut working. A release whose setup fails was rolled back to the running version, and a release signed with another key was refused. sunoto update installed a good release by hand.
Sunoto 0.1.10
Text arrives almost as soon as you tap, long dictations stay fast, the speech engine uses less memory, and the pill shows on the desktop you're on.
- Text right after the tap. Sunoto starts transcribing during the short pause before you tap to finish. If you tap without saying more, the work is already done. In testing, the wait after the tap fell from about 0.6 s to a few milliseconds, and to about 0.3 s at most with a very quick tap. The 0.1.9 engine took about 1.1 s after a 7-second dictation.
- Long dictations stay fast. A dictation over 10 seconds is transcribed in pieces at natural pauses while you talk. A one-minute dictation finishes as quickly as a short one, and reads the same as a single pass.
- Smaller memory spike. The speech engine peaks at about 2.5 GB instead of 3.3–4.3 GB, and drops back to about 1.4 GB after each dictation.
- The pill shows on the desktop you're on. With more than one desktop, the pill could open on another one while you dictated; now it moves to yours.
- A calmer pill. It no longer shows guessed words while you talk, because they were often wrong. The start hint fades once you speak, and a check mark confirms the paste.
New installs use the new speech engine. An upgrade keeps the engine already saved in your settings, and installs from before 0.1.10 saved the older one. To switch, run this and then the installer:
sed -i '' 's/"parakeet_mlx_streaming"/"parakeet_mlx_offline"/' ~/Library/Application\ Support/sunoto/config.jsonInstall or upgrade:
curl -fsSL https://raw.githubusercontent.com/shakdwipeea/voice-dictation/master/install-macos.sh | bashValidation: cargo test, clippy, and the Python suites pass. Eight dictations of 4.8 to 74 s were played to the speech engine at real-time speed. Against transcribing after the tap, the median wait after the tap fell from 571 ms to 3 ms and the worst from 974 ms to 265 ms, with the same text and memory. The pill fix was reproduced and confirmed across two desktops, and the rebuilt app installed and came up ready.
Sunoto 0.1.9
Sunoto now tells you its version, and older installs move to the Right Option shortcut.
- Right Option for everyone. Installs from 0.1.6 or earlier still had the old Ctrl+F1 shortcut saved, so upgrading kept it. Running the installer again now switches a saved Ctrl+F1 to Right Option and says so. A shortcut you picked yourself is kept, and no other setting changes.
sunoto --version. It prints the version;sunoto statusshows the version of the running app, and the installer and log name it too. Earlier releases all reported 0.1.0.- Safer uninstall.
sunoto uninstall --helpprints help instead of uninstalling, and unknown options are refused. - Shorter README with just install, use, and the everyday commands.
Install or upgrade (this also applies the shortcut change):
curl -fsSL https://raw.githubusercontent.com/shakdwipeea/voice-dictation/master/install-macos.sh | bashTo use the commands, add the alias from the README:
alias sunoto="$HOME/Applications/Sunoto.app/Contents/MacOS/sunoto-daemon"
sunoto --versionValidation: cargo test and clippy pass. An install with Ctrl+F1 saved was upgraded: setup switched it to Right Option, changed no other setting, and the app came up ready as 0.1.9.