dictate-mlx 2.2.1

Push-to-Talk Voice Dictation menu bar app for macOS using Apple Silicon MLX

Dictate

Push-to-talk voice dictation for macOS. Runs 100% on-device using Apple Silicon MLX models. No cloud, no API keys, no subscriptions.

Hold a key, speak, release — clean text appears wherever your cursor is.

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. Quit from the menu bar icon.

macOS will prompt for Accessibility and Microphone permissions on first run. Models download automatically in the background (~2-4GB total, 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+
• ~4GB RAM minimum, ~6GB recommended

How It Works

Hold PTT → Speak → Release → Clean text pasted into active window

Under the hood:

1. Push-to-talk captures audio via the microphone
2. VAD segments speech from silence
3. STT transcribes locally (Whisper or Parakeet)
4. Smart skip detects clean short phrases and skips cleanup entirely
5. LLM fixes grammar, punctuation, and formatting
6. Auto-paste puts the result wherever your cursor is

Everything runs locally. Nothing leaves your machine.

Controls

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 from the menu bar: Left Control, Right Control, Right Command, or either Option key.

STT Engines

Both engines are included. Switch anytime from the menu bar.

Engine	Speed	Languages	Notes
Parakeet TDT 0.6B	~50ms	English	Default. 4-8x faster than Whisper
Whisper Large V3 Turbo	~300ms	99+	Best for multilingual or non-English

Parakeet is the default for speed. Switch to Whisper from the menu bar if you need non-English STT.

Writing Styles

Style	What it does
Clean Up	Fixes punctuation and capitalization — keeps your words
Formal	Rewrites in a professional tone
Bullet Points	Distills your dictation into concise key points

Quality Presets

Preset	Speed	RAM	Best for
Smart	~250ms	0	Auto-routes: fast local for short, API server for long
Speedy (1.5B)	~120ms	1GB	Quick fixes, great for any chip
Fast (3B)	~250ms	2GB	Quick cleanup, everyday use
Balanced (7B)	~350ms	5GB	Longer dictation, formal rewriting
Quality (14B)	~500ms	9GB	Best accuracy for bullet points and rewrites

Times measured on M3 Ultra. The app picks the best default for your chip — Ultra/Max get 3B, everything else gets 1.5B.

The Quality menu only shows models you've downloaded. To add a larger model:

python -c "from mlx_lm import load; load('mlx-community/Qwen2.5-7B-Instruct-4bit')"

Menu Bar

All settings accessible from the waveform icon in your menu bar:

Main menu:

• Writing Style — Clean Up, Formal, or Bullet Points
• Quality — model size (shows only downloaded models)
• Input Device — select microphone
• Recent — last 10 transcriptions, click to re-paste

Advanced settings:

• STT Engine — Whisper or Parakeet
• PTT Key — choose your push-to-talk modifier
• Languages — input and output language (12 languages for translation)
• Sounds — 6 tones or silent
• LLM Endpoint — configure API server
• LLM Cleanup — toggle on/off
• Personal Dictionary — names, brands, technical terms always spelled correctly
• Launch at Login — auto-start on boot

API Server Setup

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.

Smart Routing

The Smart preset auto-routes based on message length:

• Short (15 words or fewer) → fast local model (~120ms)
• Long (16+ words) → your API server for higher quality

Environment Variables

Variable	Description	Default
DICTATE_AUDIO_DEVICE	Microphone device index	System default
DICTATE_OUTPUT_MODE	type or clipboard	type
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	qwen-1.5b, qwen, qwen-7b, qwen-14b	qwen
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.

Debugging

Run in the foreground to see logs:

dictate --foreground

Or check the background log:

tail -f ~/Library/Logs/Dictate/dictate.log

Security

• All processing is local. Audio and text never leave your machine.
• LLM endpoints are restricted to localhost by default. Set DICTATE_ALLOW_REMOTE_API=1 to override.
• Preferences stored with 0o600 permissions (owner-only read/write).
• No API keys, tokens, or accounts required.

License

MIT — See LICENSES.md for dependency licenses.