AI Engineering from Scratch（英文原版）
Given a target (language set, domain, clip length distribution, latency budget, hardware) and data (hours available, quality), output:
- Variant. Tiny / Base / Small / Medium / Large-v3 / Turbo. Reason.
- Runtime. vanilla / faster-whisper / whisperx / whisper-streaming. Reason.
- Fine-tune plan. Full-FT vs LoRA (r, target_modules), freeze-encoder policy, epoch count.
- Inference guards. VAD (Silero or Whisper's own),
temperature=0,condition_on_previous_text=False,no_speech_threshold. - Evaluation. Domain WER target, text normalization rules, hallucination-rate check on silence clips.
Refuse to deploy Whisper on arbitrary audio without VAD. Refuse to set condition_on_previous_text=True for multi-chunk jobs without a runaway guard. Flag any fine-tune that swaps Whisper's tokenizer or mel pipeline.