Local-AI-voice

Voice dictation tool that converts speech to text using AI models running locally on your machine. It is designed to work fully offline: audio is processed on-device, without sending data to external services.

This tool uses OpenVINO GenAI Whisper and takes advantage of modern hardware acceleration such as NPUs or GPUs when available, while remaining usable on standard CPUs.

The focus is on privacy, autonomy, and predictable behaviour: your audio never leaves the machine, models are locally managed, and the software can be integrated into scripts, editors, or command line workflows.

Usage

Python Version

Use Python 3.11. The Makefile defaults to py -3.11, and the audio dependencies in this repo are expected to be installed against that interpreter.

Browser streaming requires:

• websockets for uvicorn WebSocket support
• av for decoding browser Opus chunks on the server

Install dependencies

py -3.11 -m venv .venv
.venv\Scripts\activate
make install

Run desktop web UI

Open the desktop app wrapper around the local web UI:

py -3.11 ./local-ai-voice.py

The desktop window runs the same local browser UI through pywebview and starts the local transcription server automatically.

Make target:

make run

Run the raw web server without the desktop wrapper:

py -3.11 ./local-ai-voice.py --server

Make target:

make run-server

Run local transcription

Force CLI mode for file or live microphone transcription:

Transcribe a WAV file:

py -3.11 ./local-ai-voice.py --cli input.wav --model ./whisper-tiny-fp16-ov

Run live microphone transcription:

py -3.11 ./local-ai-voice.py --cli --model ./whisper-tiny-fp16-ov --chunk-seconds 1.0

Noise reduction and WebRTC VAD speech gating are enabled by default. Disable them with:

py -3.11 ./local-ai-voice.py --cli --no-silence-detect --model ./whisper-tiny-fp16-ov input.wav

Run browser transcription

Open the desktop web UI:

make run

Equivalent direct command:

py -3.11 ./local-ai-voice.py --model ./whisper-tiny-fp16-ov

The browser UI captures microphone audio in the browser and streams Opus over a WebSocket connection to the local server. The server decodes Opus, runs noise reduction and VAD at the decoded sample rate, and only resamples to 16 kHz immediately before Whisper inference.

Browser defaults:

• browser DSP is off by default:
  • Echo cancellation
  • Noise suppression
  • Auto gain control
• Voice enhance is on by default
• VAD default is 3
• overlap default is 0.00s

Notes:

• Chromium commonly uses audio/webm;codecs=opus
• Save WAV capture records the exact 16 kHz mono audio sent to Whisper
• Client debug controls browser-side and server-side debug messages shown in the page

Build standalone executable

Build the unified executable:

make build

The resulting binary supports both modes:

.\dist\local-ai-voice.exe
.\dist\local-ai-voice.exe --cli input.wav --model .\whisper-tiny-fp16-ov
.\dist\local-ai-voice.exe --server --model .\whisper-tiny-fp16-ov

License

Local-AI-voice is Copyright (C) 2026 by the Dyne.org Foundation

It is distributed under the Affero GNU General Public License v3