README.md

ASR worker microservices

The orchestrator only talks to ASR through an OpenAI-compatible POST /v1/audio/transcriptions endpoint. One worker is bundled; others are documented contracts you can implement separately.

Contract

Request (multipart/form-data):

Field	Required	Notes
file	yes	Audio chunk (16 kHz mono WAV is what the orchestrator sends)
model	yes	Model identifier
language	no	BCP-47 hint, e.g. hi, en
prompt	no	Optional decoding bias
response_format	no	verbose_json is preferred; json is acceptable
timestamp_granularities[]	no	Repeated values; the orchestrator sends segment and word

Response (JSON):

{
  "text": "haan let's start karte hain",
  "language": "hi",
  "segments": [
    {"start": 0.0, "end": 1.5, "text": "haan let's start", "confidence": 0.91}
  ],
  "words": [
    {"start": 0.0, "end": 0.4, "word": "haan", "confidence": 0.95}
  ]
}

Timestamps in the response must be local to the audio chunk; the orchestrator adds the chunk's global start_offset.

Sanity check with curl

curl -X POST http://localhost:8001/v1/audio/transcriptions \
  -F file=@chunk.wav \
  -F model=Qwen3-ASR-1.7B \
  -F language=hi \
  -F response_format=verbose_json \
  -F 'timestamp_granularities[]=segment' \
  -F 'timestamp_granularities[]=word'

Workers

• Bundled (in PyPI wheel): src/resilient_stt/workers/qwen_transformers_service/ — qwen-asr on CPU/MPS; auto-started by the orchestrator or via python scripts/bootstrap_qwen_asr_fallback.py from a git checkout.
• Documented only (under repo workers/):
  • qwen_vllm_service/ — Qwen3-ASR via vLLM (Linux + NVIDIA); bootstrap: python scripts/bootstrap_vllm_qwen3_asr.py --install
  • parakeet_openai_service/ — Placeholder: wrap NeMo/Parakeet behind the OpenAI shape
  • whisper_openai_service/ — Placeholder: wrap faster-whisper / whisper.cpp behind the OpenAI shape