Tune Whisper decoding: force English/transcribe task, beam search, repeat-ngram suppression
Destiny Aigbe
7f85a80
Greedy decoding (the previous default) was letting the model hallucinate
whole invented clauses on ambiguous/quiet audio instead of just mistranscribing
words. Beam search + no_repeat_ngram_size cuts down on both hallucination and
repetition loops; forcing language=en/task=transcribe avoids occasional
per-chunk language misdetection on short segments.
