Speech-to-text dictation for Linux. Speak into your microphone and text appears in whatever window has focus.

Supports two transcription backends:

• Local Whisper (default) — fast, private, no API key needed. Uses faster-whisper (CTranslate2 backend) for 2-4x faster inference than stock Whisper, with GPU acceleration.
• LLM via OpenRouter — uses any audio-capable model on OpenRouter (defaults to Gemini 2.0 Flash).

How it works

1. hotmic_toggle.sh starts/stops dictation via a single keybinding
2. sox records continuously from your microphone (no audio gaps)
3. Each chunk is transcribed by the configured backend (local Whisper or OpenRouter LLM)
4. The result is typed into the focused window via xdotool
5. A pulsing red "REC" overlay badge shows while recording

Audio is streamed continuously via a pipe to the worker process — no gaps between chunks. The worker splits on speech pauses or every 20s, and transcribes in a background thread while audio keeps flowing. The Whisper model stays loaded to avoid reload latency.

Requirements

• Linux with X11
• sox (audio recording and silence detection)
• xdotool
• python3 with PyGObject and cairo (for the recording indicator)

For Whisper backend (default):

• python3 with faster-whisper installed (pip install faster-whisper nvidia-cublas-cu12 nvidia-cudnn-cu12)
• CUDA GPU recommended (falls back to CPU)
• For GPU: NVIDIA CUDA libraries (pip install nvidia-cublas-cu12 nvidia-cudnn-cu12)

For LLM backend:

• curl
• jq
• An OpenRouter API key

Install dependencies (Debian/Ubuntu)

sudo apt install sox xdotool python3-gi python3-gi-cairo gir1.2-gtk-3.0

# For whisper backend:
pip install faster-whisper nvidia-cublas-cu12 nvidia-cudnn-cu12

# For LLM backend:
sudo apt install curl jq

Install dependencies (Arch)

sudo pacman -S sox xdotool python-gobject python-cairo

# For whisper backend:
pip install faster-whisper nvidia-cublas-cu12 nvidia-cudnn-cu12

# For LLM backend:
sudo pacman -S curl jq

Setup

1. Clone this repo and make the scripts executable:

git clone https://github.com/bmilleare/hotmic.git
cd hotmic
chmod +x hotmic_toggle.sh hotmic_start.sh hotmic_stop.sh hotmic_indicator.py hotmic_whisper_worker.py

2. Configure your backend

Whisper (default) — no configuration needed. Just ensure openai-whisper is installed.

LLM — set HOTMIC_BACKEND=llm and provide your OpenRouter API key:

echo 'OPENROUTER_API_KEY="sk-or-v1-your-key-here"' > /path/to/hotmic/.env

Alternatively, create ~/.config/hotmic/env with the same content, or export it in your shell profile. The script checks these locations in order:

1. Environment variable (already set)
2. .env file next to the script
3. ~/.config/hotmic/env

3. Bind hotmic_toggle.sh to a keyboard shortcut in your desktop environment's settings. For example, in GNOME:

Settings > Keyboard > Custom Shortcuts > Add:
  Name: Dictation
  Command: /path/to/hotmic/hotmic_toggle.sh
  Shortcut: (your choice, e.g. Insert)

4. Press your shortcut, speak, press it again. Text appears in the focused window.

Configuration

Edit the variables at the top of hotmic_start.sh, or override them via environment variables:

Variable	Default	Description
HOTMIC_BACKEND	whisper	whisper (local) or llm (OpenRouter)
WHISPER_MODEL	medium.en	Whisper model: tiny, base, small, medium.en, large-v3-turbo, etc.
WHISPER_DEVICE	cuda	cuda for GPU, cpu for CPU-only
OPENROUTER_MODEL	google/gemini-2.0-flash-001	Any audio-capable model on OpenRouter (LLM backend only)
SILENCE_START_THRESH	3%	Threshold to detect speech start (must be above ambient noise)
SILENCE_STOP_THRESH	3%	Threshold to detect pauses (must be above ambient noise)
SILENCE_DUR	0.8	Seconds of silence before a chunk ends
MAX_CHUNK_SEC	20	Hard cap per chunk (silence split handles most cases)

Files

File	Purpose
hotmic_toggle.sh	Start/stop dictation (bind to a hotkey)
hotmic_start.sh	Main recording + transcription loop
hotmic_stop.sh	Stops recording and cleans up
hotmic_indicator.py	Pulsing "REC" overlay badge
hotmic_whisper_worker.py	Persistent Whisper process (loads model once)

Logs

Logs are written to /tmp/hotmic/hotmic.log for debugging.

License

MIT