# WebRTC dictation

This Bazel package runs a local WebRTC speech-to-text service using aiortc and
faster-whisper. The default multilingual Whisper `small` model uses the RTX
4070 Ti through native WSL CUDA with `int8_float16` compute.

## Setup

```bash
bazel run //dictation:preflight
bazel run //dictation:download_model
bazel run //dictation:server
```

Open <http://127.0.0.1:8090>, grant microphone permission, and start
dictation. Audio is carried by WebRTC and partial/final transcript events are
returned on the `transcripts` RTC data channel.

To transcribe an existing audio file:

```bash
bazel run //dictation:transcribe -- /path/to/audio.wav
```

To exercise the running HTTP signaling and WebRTC path with an audio file:

```bash
bazel run //dictation:webrtc_smoke -- /path/to/audio.wav
```

The model is stored under `~/.cache/zenbu/faster-whisper-small` by default.
Override it with `DICTATION_MODEL_DIR`.

Useful tuning:

```bash
DICTATION_COMPUTE_TYPE=int8_float16 \
DICTATION_MAX_SESSIONS=1 \
DICTATION_PARTIAL_INTERVAL_MS=500 \
DICTATION_SILENCE_MS=400 \
  bazel run //dictation:server
```

The first version intentionally binds to loopback and does not configure STUN
or TURN. Remote WebRTC access requires HTTPS, authentication, origin policy,
and usually a TURN service.

Infinite Canvas hosts this page as a hidden CEF/iframe transport. Pressing `M`
shows partial and final transcript text in a centered Raylib caption; pressing
`Enter` sends the accumulated text to Qwen session orchestration:

```bash
bazel run //infinite_canvas:agent_dev
```