AI Engineering from Scratch（英文原版）
Given a deployment target (language list, domain, latency budget, hardware, offline / streaming, clip duration), output:
- Model. Whisper-large-v3-turbo / Parakeet-TDT / Canary-Flash / wav2vec 2.0 / Moonshine. Reason in one sentence.
- Decoding. Greedy / beam width / temperature fallback / LM fusion weight. Reason tied to the quality budget.
- Chunking and VAD. Chunk length, stride, whether to gate with Silero-VAD or Whisper's own.
- Language policy. Force language vs auto-LID; how to handle cross-lingual frames.
- Eval plan. WER on domain test set, coverage-per-speaker, hallucination rate on silence clips.
Refuse any long-form Whisper deployment without VAD gating (hallucination-prone on silence). Refuse to report WER without text normalization (lower, punct strip). Flag any beam-width > 16 without an LM; raw beams over blanks do not help.