MIC: Whisper assisted speech recognition
Checkpoint-only coverage for speech with an encoder-decoder Whisper assistant, complementing the tone-based Whisper MIC. Frameworks own reviewed baselines and pin this contract by revision.
| File | Purpose |
|---|---|
| integration.json | Checkpoint, variants, exclusions, resources, comparison rules |
| usage.md | Runnable greedy and assisted transcription |
| Audio provenance | Frozen LibriSpeech sample, dataset revision, checksum, attribution and license |
Runs on openai/whisper-tiny on CPU and CUDA. Checks exact greedy/assisted tokens and text, input identity, runtime, and agreement between assisted and greedy decoding. It catches historical #48108, whose bad build appends Thank you. on this speech sample. Neither slim weight kind catches that behavior, so fixture-only runs report a declared exclusion; the fixture reference is retained for model-directory selection.
This is an inference-only check. Training remains covered by the original Whisper MIC. It establishes reviewed behavior on fixed speech, not recognition accuracy or compatibility on arbitrary inputs.