dictate-mlx 2.5.0

65ms voice-to-text on Apple's Neural Engine. Push-to-talk dictation for macOS — 100% local, free, open source.

Push-to-talk voice dictation that runs entirely on your Mac.
No cloud. No API keys. No subscriptions.

Hold a key → Speak → Release → Clean text appears wherever your cursor is.

---

Why Dictate?

• 65ms voice-to-text on Apple's Neural Engine — faster than a keystroke
• Zero GPU RAM for STT — the Neural Engine has its own dedicated memory
• 100% local — audio and text never leave your Mac
• Free and open source — no subscriptions, no API keys, no accounts
• LLM text cleanup — local model fixes grammar and punctuation automatically
• 12 languages — real-time translation between any supported pair

Your M-series Mac has a 16-core Neural Engine doing nothing. Dictate puts it to work.

Install

pip install dictate-mlx
dictate

That's it. Dictate launches in the background and appears in your menu bar. Close the terminal — it keeps running.

macOS will prompt for Accessibility and Microphone permissions on first run. Models download automatically (~1-3GB depending on preset, cached in ~/.cache/huggingface/).

Install from source

git clone https://github.com/0xbrando/dictate.git
cd dictate
python3 -m venv .venv
source .venv/bin/activate
pip install -e .
dictate

Requirements

• macOS with Apple Silicon (any M-series chip)
• Python 3.11+
• ~3GB RAM with ANE (STT runs on Neural Engine, only LLM needs GPU memory)

Features

🎙️ Push-to-Talk

Hold a key, speak, release. Text appears wherever your cursor is.

Action	Key
Record	Hold Left Control
Lock recording (hands-free)	Press Space while holding PTT
Stop locked recording	Press PTT again

The PTT key is configurable: Left Control, Right Control, Right Command, or either Option key.

🧠 LLM Text Cleanup

The thing that sets Dictate apart. Most dictation tools give you raw transcription. Dictate pipes through a local LLM that fixes grammar, adds punctuation, and formats properly.

Short phrases (≤15 words) skip cleanup for instant speed. Longer dictation gets the full treatment.

🗣️ Three STT Engines

All included. Switch anytime from the menu bar.

Engine	Speed	Languages	Notes
ANE (Neural Engine)	~65ms	25	Default — runs on Apple Neural Engine, frees GPU for LLM
Parakeet TDT 0.6B	~50ms	25	Runs on GPU via MLX
Whisper Large V3 Turbo	~300ms	99+	Auto-selected for Japanese, Chinese, Korean

ANE is the default. It runs speech recognition on Apple's Neural Engine — a dedicated chip that sits idle during most tasks. This frees the GPU entirely for LLM text cleanup, so STT and LLM run concurrently with zero contention. The result: 65-106ms transcription on real speech.

When you select a language not supported by ANE/Parakeet (Japanese, Chinese, Korean), Dictate automatically switches to Whisper. Switch back to a supported language and it returns to ANE.

✍️ Writing Styles

Style	What it does
Clean Up	Fixes punctuation and capitalization — keeps your words
Professional	Polished tone and grammar
Bullet Points	Rewrites as concise bullet points

Toggle LLM cleanup off from the menu bar for raw transcription output.

🌐 Real-Time Translation

Speak in one language, get output in another. 12 languages supported: English, Spanish, French, German, Italian, Portuguese, Japanese, Korean, Chinese, Russian, Arabic, Hindi.

⚡ Quality Presets

Preset	Speed	Size	Best for
Qwen2.5 1.5B	~250ms	950MB	Fast and reliable (default)
Qwen2.5 3B	~400ms	1.8GB	Best accuracy
API Server	varies	0	Use your own LLM server (LM Studio, Ollama, etc.)

Short phrases (15 words or less) skip LLM cleanup entirely for instant output. The app picks the best default model for your chip.

End-to-End Pipeline

Full latency from voice → text on screen:

Mode	GPU RAM	Latency
LLM off (raw transcription)	0	~65ms
LLM on (Qwen2.5 1.5B)	~950MB	~315ms
LLM on (Qwen2.5 3B)	~1.8GB	~465ms

With ANE, speech recognition runs on a dedicated chip with its own memory — zero GPU usage. Turn off LLM cleanup and the entire app uses no GPU RAM at all.

Menu Bar

Everything accessible from the waveform icon:

• Writing Style — Clean Up, Professional, Bullet Points
• Quality — Qwen2.5 1.5B or 3B (or API server)
• Input Device — select microphone
• Recent — last 10 transcriptions, click to re-paste
• STT Engine — ANE (default), Parakeet, or Whisper
• PTT Key — choose your push-to-talk modifier
• Languages — input and output language
• Sounds — 6 notification tones or silent
• Personal Dictionary — names, brands, technical terms always spelled correctly
• Launch at Login — auto-start on boot

ANE Engine Setup

The ANE (Apple Neural Engine) engine is the default and recommended STT engine. It requires a small Swift binary that Dictate calls behind the scenes. If the binary isn't installed, Dictate falls back to Parakeet (GPU-based STT).

# Build from source (requires Xcode command line tools)
cd swift-stt
swift build -c release

# The binary lands at swift-stt/.build/release/dictate-stt
# Either add it to your PATH or leave it — Dictate finds it automatically

First run: CoreML models download automatically (~2.7GB) and compile for your chip. This takes 1-2 minutes the first time. After that, models are cached and transcription starts instantly.

Requirements: macOS 14+ (Sonoma or later), Apple Silicon.

What it does: The dictate-stt binary uses FluidAudio to run Parakeet speech recognition on the Neural Engine via CoreML. All processing is local — no network calls after the initial model download.

How it works

When you select ANE in the menu bar, Dictate calls the dictate-stt binary as a subprocess:

1. Dictate records audio and saves it as a temporary WAV file
2. Calls dictate-stt transcribe /tmp/audio.wav
3. The Swift binary runs the audio through CoreML on the Neural Engine
4. Returns JSON to stdout: {"text": "Hello world", "duration_ms": 68}
5. Dictate parses the result and pipes it through LLM cleanup as usual

The binary is a standalone executable with no Python dependency. You can also use it directly:

dictate-stt check                    # Verify ANE is available
dictate-stt transcribe recording.wav # Transcribe a WAV file

API Server

If you run a local LLM server, Dictate can use it instead of loading its own model — zero additional RAM:

DICTATE_LLM_BACKEND=api DICTATE_LLM_API_URL=http://localhost:8005/v1/chat/completions dictate

Works with any OpenAI-compatible server: vllm-mlx, LM Studio, Ollama.

The Smart preset auto-routes by length: short phrases → fast local model (~120ms), longer dictation → your API server.

Environment Variables

All environment variables

Variable	Description	Default
DICTATE_AUDIO_DEVICE	Microphone device index	System default
DICTATE_OUTPUT_MODE	type or clipboard	type
DICTATE_STT_ENGINE	ane, parakeet, or whisper	ane
DICTATE_INPUT_LANGUAGE	auto, en, ja, ko, etc.	auto
DICTATE_OUTPUT_LANGUAGE	Translation target (auto = same)	auto
DICTATE_LLM_CLEANUP	Enable LLM text cleanup	true
DICTATE_LLM_MODEL	qwen25-1.5b, qwen-3b	qwen25-1.5b
DICTATE_LLM_BACKEND	local or api	local
DICTATE_LLM_API_URL	OpenAI-compatible endpoint	http://localhost:8005/v1/chat/completions
DICTATE_ALLOW_REMOTE_API	Allow non-localhost API URLs	unset

Agent Integration

Dictate works well as a voice input layer for AI assistants and agent frameworks. If you're building with tools like Claude Code, OpenClaw, or similar — Dictate gives your setup a local, private voice interface with zero cloud dependency.

CLI Commands

dictate              # Launch in menu bar (backgrounds automatically)
dictate config       # View all preferences
dictate config set writing_style professional
dictate config set quality fast
dictate config set ptt_key cmd_r
dictate config set stt whisper
dictate config reset # Reset to defaults
dictate stats        # Show usage statistics
dictate status       # System info and model status
dictate doctor       # Run diagnostic checks (troubleshooting)
dictate devices      # List audio input devices
dictate update       # Update to latest version
dictate -f           # Run in foreground (debug)
dictate -V           # Show version

Config Keys

Key	Values
writing_style	clean, professional, bullets
quality	api, fast, quality
stt	ane, parakeet, whisper
input_language	auto, en, ja, de, fr, es, ...
output_language	auto, en, ja, de, fr, es, ...
ptt_key	ctrl_l, ctrl_r, cmd_r, alt_l, alt_r
llm_cleanup	on, off
sound	soft_pop, chime, warm, click, marimba, simple
llm_endpoint	host:port (for API backend)

Shell Completions

Tab completions for bash and zsh:

# Bash — add to ~/.bashrc
source /path/to/dictate/completions/dictate.bash

# Zsh — copy to fpath dir, then reload
cp completions/dictate.zsh ~/.zsh/completions/_dictate
autoload -Uz compinit && compinit

Completes commands, config keys, and all valid values.

Debugging

# Run in foreground with logs
dictate --foreground

# Check background logs
tail -f ~/Library/Logs/Dictate/dictate.log

Security

• All processing is local. Audio and text never leave your machine.
• The ANE engine's dictate-stt binary is open source Swift code you build yourself from swift-stt/. CoreML models download from Hugging Face on first run, then everything is cached locally.
• LLM endpoints restricted to localhost by default (DICTATE_ALLOW_REMOTE_API=1 to override).
• Preferences stored with 0o600 permissions (owner-only).
• No API keys, tokens, or accounts required.

Contributing

Issues and PRs welcome. Run the test suite before submitting:

python -m pytest tests/ -q

See CONTRIBUTING.md for guidelines.

License

MIT — See LICENSES.md for dependency licenses.