| name | ambient-qwen3-asr |
| description | Local Qwen3-ASR speech-to-text provider for Ambient push-to-talk. |
Ambient Qwen3-ASR STT Provider
Use this package only through Ambient's STT provider path. It implements:
qwen3_asr_transcribe --audio <wav> --language <language> --output-json <path>
The command expects Ambient to provide a normalized 16 kHz mono WAV. It writes concise JSON with text, language, durationMs, providerId, and artifact metadata. It must not print raw audio, secrets, or long diagnostic payloads to stdout.
Runtime configuration:
AMBIENT_QWEN3_ASR_BINARY: optional path or command name for llama-mtmd-cli.
- Default model assets are pinned in
assets/qwen3-asr-assets.json, downloaded into an Ambient cache on first use, verified by exact byte size and SHA-256, then passed to llama.cpp with -m and --mmproj.
AMBIENT_QWEN3_ASR_ASSET_DIR: optional cache root for those verified manifest assets.
AMBIENT_QWEN3_ASR_MODEL: optional custom llama.cpp -hf model ref for explicit experiments. Setting it bypasses the default verified manifest asset path.
AMBIENT_QWEN3_ASR_THREADS: optional thread count. Defaults to a conservative local value.
AMBIENT_QWEN3_ASR_FAKE_TRANSCRIPT: deterministic test hook. Do not set this for product use.
Product constraints:
- Use the exact transcript as a visible user message.
- Do not use partial text to steer an active agent request.
- Let Ambient's no-speech gate run before invoking this provider.
- Do not enable automatic runtime archive downloads until every per-platform runtime lane has a pinned source, size, and SHA-256. Windows remains pending real-machine validation.