whisper: align no_speech_threshold fallback condition with openai/whisper

Summary

The MLX Whisper decode_with_fallback logic appears to differ from upstream openai/whisper in how no_speech_threshold suppresses fallback decoding.

In upstream OpenAI Whisper, high no_speech_prob only disables fallback when the average log probability is also below logprob_threshold:

if (
    no_speech_threshold is not None
    and decode_result.no_speech_prob > no_speech_threshold
    and logprob_threshold is not None
    and decode_result.avg_logprob < logprob_threshold
):
    needs_fallback = False  # silence

In mlx_whisper.transcribe, the condition is broader:

if (
    no_speech_threshold is not None
    and decode_result.no_speech_prob > no_speech_threshold
):
    needs_fallback = False  # silence

Why this matters

This means MLX Whisper can accept a decode as “silence” solely because no_speech_prob is high, even if fallback was triggered for another reason such as high compression ratio / repetitive output.

The upstream behavior is narrower: it only suppresses fallback when the segment looks like silence according to both high no-speech probability and low average log probability. That may affect hallucination/repetition behavior.

References

OpenAI Whisper source: https://github.com/openai/whisper/blob/main/whisper/transcribe.py

Relevant upstream condition is in decode_with_fallback.

MLX Whisper source: whisper/transcribe.py, inside decode_with_fallback.

Suggested change

Match upstream OpenAI Whisper by adding the logprob_threshold and avg_logprob checks:

if (
    no_speech_threshold is not None
    and decode_result.no_speech_prob > no_speech_threshold
    and logprob_threshold is not None
    and decode_result.avg_logprob < logprob_threshold
):
    needs_fallback = False  # silence