Skip to content
olchuPublic

About

Local voice dictation for macOS: hold Fn, speak, release — text is pasted where your cursor is. On-device Whisper & Parakeet, no cloud.

Topics

Resources

Stars

1 star

Watchers

0 watching

Forks

Repository files navigation

FlowType, voice dictation that runs on your Mac

FlowType

macOS 14+ Swift Apple Silicon License: MIT with exceptions

FlowType is a free, open-source menu bar app that lets you dictate into any app on your Mac. Hold Fn, speak, let go, and the text is pasted wherever your cursor is.

Recognition runs on your Mac, with Whisper, NVIDIA Parakeet or Sber GigaAM. Your audio stays in memory and is never written to disk, and the app needs the network only once, to download the models. If you want a stronger model to clean up the text, you can connect your own, in the cloud or on your Mac.

How it works

Hold Fn, speak, release, and the text is pasted

  1. Hold Fn in any app. A short cue plays and the indicator appears above the Dock.
  2. Speak. The indicator follows your voice, and Whisper models start transcribing while you're still talking.
  3. Release Fn. FlowType trims the silence, decodes the remaining tail and cleans up the text.
  4. The text is pasted into the field that had focus, and your clipboard gets its previous contents back.

If you press Fn and say nothing, nothing is pasted: the speech detector drops empty recordings.

To cancel, press Esc while recording or while the text is being processed. Nothing is pasted, and the Esc doesn't reach the app you're typing in.

For a long dictation, double-press Fn and let go: the recording goes on without holding the key, until you press it once more. Esc cancels it as usual. A recording you forgot about finishes by itself after 10 minutes and its text is pasted; you can change the limit in Settings → General.

If Fn doesn't suit you, Settings → General offers the right ⌥, ⌘ or ⌃ instead. Pressing another key or a mouse button along with one of them counts as a shortcut, so right ⌘ + C still copies: a recording you're holding is dropped, and a hands-free one just goes on.

If the text went nowhere, say because the cursor wasn't in a text field, put the cursor where it belongs and press ⌃⌥V. FlowType pastes the last dictation again, whatever is on the clipboard by then.

To take back what you just dictated, hold Fn and say “отмени”, “удали последнее” or “undo that”. It works right after the dictation, in the same field, while the cursor is still after it.

Each dictation fits into the text around the cursor. After a word or a comma it goes on with a small letter and a space, at the start of a line or a sentence it takes a capital, and with text after the cursor it leaves out its full stop. Names keep their capital. FlowType reads the text around the cursor through Accessibility; where an app doesn't tell it, as many Electron apps don't, a dictation right after the last one goes on from it.

Features

Speech recognition

Profile Model When to use
Fast Whisper Tiny Quick start, lower accuracy
Balanced (default) Whisper Large v3 Turbo, 632 MB Good accuracy at a small size
Accurate Whisper Large v3, 947 MB Best Whisper accuracy, slower start
Parakeet NVIDIA Parakeet TDT Ultra, 630 MB Fastest; accurate on Russian, English and 23 more European languages
GigaAM Sber GigaAM v3 (end-to-end CTC), 420 MB Most accurate on Russian, writes punctuation and numbers itself; Russian only

The language can be set to auto-detect, Russian or English.

Whisper recognises speech while you're still recording, so there's very little left to wait for once you release Fn. You choose how the final text is produced:

  • Stream + tail keeps the confirmed streaming text and re-decodes only the tail.
  • Full recording transcribes the whole recording after you release the key. It's the baseline the other modes are compared against.
  • Stream only uses the streaming text as is. It's the fastest mode, but the last words are less reliable.

Parakeet and GigaAM run on the Neural Engine and decode the whole recording in a fraction of a second. They have no streaming preview. GigaAM takes 25 seconds at a time, so longer dictations are split at the quietest moments.

Silero VAD checks whether a recording contains speech and trims the silence. Without that check, Whisper tends to hallucinate subtitle lines like "To be continued..." on silent audio.

The model is warmed up in advance and stays in memory, with loading progress shown in the menu and in Settings.

Text cleanup

The default mode, Local rules, is a set of fast built-in rules that remove repeated fragments and typical Whisper artefacts. Off leaves only basic cleanup.

Settings → Text Rules holds rules applied after recognition in every mode, each with its own switch. Remove filler words takes out words from an editable list (“ну”, “типа”, “как бы”, “короче”, “like”, “you know”, “I mean” and others). A word goes only where it stands apart, at the start of a phrase or between commas, so “Ну, короче, давай” becomes “Давай” while “что-то типа этого” stays. Hesitation sounds like “э”, “ммм”, “uh” and “um” always go.

Write numbers as digits works in Russian and English. Russian numbers turn into digits in any case: “двадцати пяти” becomes 25, “двадцать первый” becomes 21-й, “двадцать пять процентов” becomes 25%, “пять тысяч рублей” becomes 5000 ₽ and “в девять тридцать” becomes в 9:30. Dates and years are written as in edited text, “25 октября” and “в 2026 году”. In English, “twenty-five” becomes 25, “twenty-first” becomes 21st, “five dollars” becomes $5, “at nine thirty” becomes at 9:30 and “nineteen ninety” becomes 1990. Numbers under ten stay words unless you switch that off, so “у меня два кота” and “two cats” are left alone.

Write spoken punctuation turns “точка”, “запятая”, “точка с запятой”, “двоеточие”, “многоточие”, “тире”, “вопросительный знак” or “знак вопроса”, “восклицательный знак”, “открыть кавычки”, “закрыть кавычки”, “открыть скобку”, “закрыть скобку”, “новая строка” and “новый абзац”, or “period”, “comma”, “semicolon”, “colon”, “ellipsis”, “dash”, “question mark”, “exclamation mark”, “open quote”, “close quote”, “open bracket”, “close bracket”, “new line” and “new paragraph”, into marks and line breaks. Commands that are also everyday words, like “точка”, “запятая” or “новая строка”, work only where a phrase ends: before a short pause, another command or the end of the dictation. So “Привет, запятая, как дела, вопросительный знак” becomes “Привет, как дела?” while “точка зрения” stays. Names that mean nothing but a mark, like “тире” or “открыть скобку”, work anywhere: “он сказал открыть кавычки привет закрыть кавычки” becomes “он сказал «привет»”.

End without a full stop leaves out the full stop at the end of a dictation, which suits chats and continuing a sentence. Say “точка” or “period” when you want one.

The Dictionary holds words and names the recogniser gets wrong, written your way: “флоу тайп” becomes “Flow Type”. Case and “ё” don't matter, and the punctuation around the phrase stays.

Settings → Snippets holds text that a short phrase stands for: say “моя почта” on its own, and your address is pasted exactly as written. A snippet can have several phrases (“моя почта”, “мой имейл”) and can also work inside a longer phrase. Snippets are kept in the Keychain.

Local LLM applies the same rules and then runs a small language model on your Mac, built on MLX, to fix punctuation and obvious recognition errors. If its edit adds or drops more than a stray word, FlowType throws it out and keeps the rule-based text.

Model Size Notes
Qwen2.5 1.5B 0.9 GB Fastest, best for short phrases
Qwen3.5 2B 1.7 GB Noticeably better with Russian
Qwen3 4B 2.3 GB Most careful edits, about twice as slow
Gemma 4 E2B 3.6 GB Strong multilingual model from Google

External LLM applies the same rules and then sends the text to a model you connect. Besides punctuation and misheard words, it takes out what a small model can't be trusted with: false starts and slips you corrected, so “давай в среду, нет, в четверг” becomes “Давай в четверг”. Its edits go through the same check, which lets it drop more words than the local model, and an edit that rewrites the dictation is thrown out. If the model doesn't answer within 8 seconds, or 30 for a server on your Mac, which may have to load the model first, the rule-based text is pasted. Esc cancels the request along with the dictation.

Provider Default model Runs
OpenAI gpt-5.4-mini Cloud
Anthropic claude-haiku-5-5 Cloud
Google Gemini gemini-3.1-flash-lite Cloud
Groq openai/gpt-oss-20b Cloud
OpenRouter google/gemini-3.1-flash-lite Cloud
Ollama, LM Studio any model you have On your Mac
Custom any model Any server with the OpenAI chat completions API

The API key is kept in the Keychain, separately for each provider. Check lists the server's models and cleans up a test phrase, so you see that the connection works and how fast it answers. You can add your own instructions to the built-in prompt, such as “Keep English terms in Latin script”. The words from your dictionary go into the prompt too, so the model keeps their spelling.

With a cloud provider, the text of every dictation goes to that provider, and its privacy terms apply. The audio and your snippets never leave your Mac.

Pasting

FlowType puts the text on the clipboard, pastes it with ⌘V into the app where you started dictating, then restores your previous clipboard (you can turn that off). The clipboard is restored once that app has actually read the text, not after a fixed delay, so a slow app never gets your old clipboard pasted instead. If you copy something in the meantime, your copy is kept. The transcript is marked as transient, so clipboard managers like Maccy, Paste and Raycast don't save it to their history. If ⌘V fails, the text is inserted through the Accessibility API, and if that fails too, it stays on the clipboard.

Parallels Desktop gets its own handling, because it syncs the clipboard with the guest OS asynchronously. With auto-paste off, the text is only copied.

Recording indicator

Waves, Dots, Particles and Aurora playing a dictation, and the Bars pill

Waves is a set of glowing waves that open out of an orb and react to your voice. While the text is being processed, a pulse of light travels along them, and they flash when it's pasted. They come in five colours: blue, red, yellow, green and violet. The waves are drawn with Core Animation instead of SwiftUI's Metal renderer, which keeps their GPU memory use close to zero.

Bars is a compact pill with level bars. It shows a pulsing ellipsis while transcribing and a checkmark after pasting.

Dots is a field of glowing dots in columns, mirrored around a centre row like a spectrum. The columns bounce with your voice and turn white towards the middle, a bump of light sweeps across them while the text is processed, and they flash and fold in when it's pasted. It uses the same five colours as Waves.

Aurora is a glass bubble with coloured smoke turning inside it: orange, violet and blue, each winding round its own axis and lit from the upper left. Your voice lights the smoke up, brighter, hotter and a little thicker the louder you speak, stirs it faster and swells the bubble, which pulses and wobbles like jelly while you talk. While the text is processed it thinks: it shrinks, breathes evenly and whirls in its palette's thinking colours. There are four palettes: Nebula, Ocean, Ember and Borealis. It appears out of smoke gathering into a small clot that swells into the bubble, and at the end the bubble bursts and the smoke tears apart and dissolves, differently every time. It's drawn by a small Metal renderer, which keeps its memory use to a few megabytes.

Particles is a field of glowing dots seen in perspective, swelling slowly like a small landscape, with a mountain for each frequency band of your voice: the lows on the left, the highs on the right, so a vowel and an "s" raise different peaks. Its colour runs from violet and pink through blue and cyan to nearly white, the peaks turn pale, and fine dust drifts over it. Like the waves, it gathers out of the middle and folds back into it. It's drawn with Metal too, and the spectrum comes from an FFT of the microphone in Accelerate.

Every style appears almost instantly and collapses right after the paste. Microphone sensitivity is adjustable, and at low values the indicator ignores background noise.

Sounds

Short cues play when recording starts and stops and when the text is pasted, so you know where you are without glancing at the indicator. You can adjust the volume or turn sounds off, and preview each cue with its play button.

The indicator doesn't mistake the start cue for your voice. With sounds off, the waves react to speech straight away.

History

FlowType keeps the last 100 dictations from the past week. History… in the menu opens them with a search field; a click pastes one into the app in front. The history is stored on disk encrypted with AES-GCM, with the key in the Keychain, and can be cleared or switched off in Settings → General. Audio is never saved.

Menu and settings

The Settings window with the Models and Indicator panes

The menu shows the current status and model loading progress, switches the indicator style and opens the history. Settings live in one window with a sidebar, the way System Settings does it:

  • General: hotkey and double press for hands-free, microphone, auto-paste and clipboard restore, history, open at login, reset settings. The microphone follows the system's unless you choose one; a chosen one that isn't connected gives way to the system's until it's back.
  • Permissions: microphone and Accessibility. Changes you make in System Settings are picked up automatically.
  • Models: profile choice, model download and deletion, loading progress of the active model.
  • Transcription: language, streaming mode and text cleanup, including the provider, key and model of an external LLM.
  • Text Rules: fitting the text around the cursor, the full stop at the end, filler words, numbers as digits, spoken punctuation, the dictionary and the undo command.
  • Snippets: phrases that stand for longer text, such as your email or address.
  • Indicator: style (Bars, Waves, Dots, Aurora or Particles), colour or palette, position and sensitivity, with a live preview. Choose… under Position shows the indicator so you can drag it anywhere on any screen; it appears there from then on.
  • Sounds: on/off, volume and cue previews.

On first launch, an onboarding window offers to download the selected model.

Under the hood

Record, detect speech, transcribe, clean up, paste

Stage Technology
Global hotkey CGEvent tap (listen-only)
Recording AVAudioEngine, mono PCM in memory, resampled to 16 kHz
Speech detection Silero VAD via FluidAudio
Recognition WhisperKit (Core ML), NVIDIA Parakeet TDT via FluidAudio, Sber GigaAM v3 on Core ML
Cleanup Custom rules, LLM via MLX Swift LM, or an external one through the OpenAI chat completions or Anthropic Messages API
Pasting NSPasteboard, CGEvent for ⌘V, Accessibility API
UI SwiftUI MenuBarExtra, NSPanel for the indicator, Core Animation
FlowType/
  App/        AppState, indicator and onboarding controllers
  Models/     settings, profiles, models, statuses
  Services/   hotkey, recording, VAD, transcription, cleanup, pasting, permissions
  UI/         menu, settings, onboarding, indicators
  Utils/      clipboard and text helpers
  Sounds/     feedback sounds

Requirements

  • macOS 14.0 or later
  • Apple Silicon Mac (recommended)
  • Xcode 16 or later to build
  • Microphone permission for recording, and Accessibility for the global hotkey and pasting

Build and run

git clone git@github.com:olchu/flowType.git
cd flowType
open FlowType.xcodeproj

Xcode resolves the Swift packages automatically. Run the FlowType scheme, then on first launch:

  1. Grant Microphone and Accessibility access (Settings → Permissions).
  2. Download a model in the onboarding window or in Settings → Models.
  3. Put the cursor in any text field, hold Fn and speak.

If Fn opens the emoji picker or switches the input source, set "Press 🌐 key to" to Do Nothing in System Settings → Keyboard.

Building a DMG

scripts/build_dmg.sh

The script builds Release and puts the image at dist/Flow Type-<version>.dmg.

The build is signed with an Apple Development certificate and isn't notarized, so Gatekeeper blocks it the first time it's opened on another Mac. Open the app with right-click → Open, or click Open Anyway in System Settings → Privacy & Security.

Debugging text cleanup

You can run LLM cleanup from the command line without recording anything:

# warm up (and download if needed) the cleanup model
"Flow Type.app/Contents/MacOS/Flow Type" --prewarm-cleanup-model --cleanup-model "Qwen3 4B"

# clean up a string and print the result, the decision and the timing
"Flow Type.app/Contents/MacOS/Flow Type" --cleanup-transcript "so um I wanted to to say that" --cleanup-model "Qwen3.5 2B"

# the same through the external LLM from the settings
"Flow Type.app/Contents/MacOS/Flow Type" --cleanup-transcript "давай в среду нет в четверг" --cleanup-mode "External LLM"

Updating the README images

The indicator animation and the settings picture are made by scripts, so they can be redone after a change to the look:

# records Waves, Dots, Particles and Aurora off screen and builds Docs/images/indicators.webp
scripts/record_indicators.sh

# screenshots two settings panes of the installed app and builds Docs/images/settings.png
scripts/capture_settings.sh

record_indicators.sh uses --record-indicator, which plays a made-up voice through one dictation at fixed steps of time and writes every frame as a PNG; see FlowType/App/IndicatorRecorder.swift. Both scripts need Pillow, and the first one also needs img2webp from brew install webp.

Tuning the waves look

The Wave Tuner is a hidden editor for design work. Open Settings → Indicator and press ⌥ three times quickly, and a Fine-tune… button appears. In the editor you can switch each layer on or off (halo, glow, lines, orb, additive blending) and adjust its parameters while you watch the result on white, light gray and dark backgrounds. Your values are saved and applied to the real indicator. The defaults live in FlowType/UI/WaveAppearance.swift.

Privacy

Audio is processed in memory and never saved. Recognition and cleanup run locally, with no cloud APIs. The network is used only to download models from Hugging Face, unless you connect a cloud model for text cleanup: then the text of each dictation, never the audio, goes to the provider you chose.

Acknowledgements

FlowType is built on these open-source projects:

Speech and language models aren't bundled with the app. They're downloaded from Hugging Face on demand and are covered by their own licenses: OpenAI Whisper, NVIDIA Parakeet, Sber GigaAM (MIT, Core ML conversion by Ivan Ushakov), Silero VAD, Qwen and Gemma.

License

FlowType is free. Its source code is released under the MIT License, with the exceptions below.

The design of the recording indicators is not covered by the MIT License, and all rights to it are reserved. That means the look and motion of the Waves, Dots, Aurora, Particles and Bars styles, with their colours and palettes, and the files that draw them, listed in LICENSE. You may build and run FlowType from this repository for your own use, but you may not copy the indicators or reuse them in another app or product without written permission.

The FlowType name, the app icon, the menu bar icon and the images in Docs/images are not covered by the MIT License either, and all rights to them are reserved. If you publish a fork or a derived app, give it a different name and icon, and remove the indicators or replace them with your own.

About

Local voice dictation for macOS: hold Fn, speak, release — text is pasted where your cursor is. On-device Whisper & Parakeet, no cloud.

Topics

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages