Bias Whisper transcription to English with beam/ VAD tuning and auto-language fallback

[Gokias]

Motivation

• Improve transcription accuracy for speakers where the model drifts or guesses the wrong language by biasing toward English and strengthening decode settings.
• Provide configurable knobs to tune accuracy vs CPU cost and to fall back to auto language when forced-English output is low-confidence.

Description

• Add new environment-configurable settings: TRANSCRIBE_LANGUAGE (default en), TRANSCRIBE_BEAM_SIZE (default 5), and TRANSCRIBE_FALLBACK_AVG_LOGPROB (default -1.2).
• For faster_whisper, set language=TRANSCRIBE_LANGUAGE and stronger decode options (task="transcribe", condition_on_previous_text, beam_size, and vad_parameters) and centralize segment collection in a _collect_segments helper.
• Implement a two-pass fallback for faster_whisper that retries with language=None (auto-detect) when the average log-prob of the forced-English pass is below the configured threshold and selects the fallback only if its average log-prob is better.
• For the whisper engine path, bias calls by passing language=TRANSCRIBE_LANGUAGE and task="transcribe".

Testing

• Ran python -m py_compile poopbot.py to validate syntax and it completed successfully.

---

Codex Task