Idea: postprocess results with high compression ratio

[dgoryeo]

Hi @jianfch ,

I came across this PR suggestion in the main Whisper community: Add compression_ratio_hallucination_threshold.
I thought may be stable-ts can achieve this as part of its enhancement/refinement? For example to redo those segments, or discard them if they're above the threshold.

Just a thought.

Thanks

[jianfch]

The redo is already handled by compression_ratio_threshold.

stable-ts/stable_whisper/whisper_word_level/original_whisper.py

Lines 99 to 100 in 4711a01

	compression_ratio_threshold : float, default 2.4
	If the gzip compression ratio is above this value, treat as failed.

stable-ts/stable_whisper/whisper_word_level/original_whisper.py

Lines 352 to 356 in 4711a01

	if (
	compression_ratio_threshold is not None
	and decode_result.compression_ratio > compression_ratio_threshold
	):
	needs_fallback = True # too repetitive

Currently, there's no parameter that directly removes those segments but you can do something like this:

for seg in reversed(result):
    if seg.compression_ratio > 3.0:
        result.remove_segment(seg)

[jianfch]

After 08421e2 this can be implemented with one line or just an argument:

result.custom_operation('compression_ratio', '>', 3.0, 'remove', False)
# or
result.regroup('co=compression ratio+>+3.0+remove+0')

Those segments should be removed before other regroup methods move things around:

result = model.transcribe(..., regroup='co=compression ratio+>+3.0+remove+0')
# or to follow it up with default regrouping
result = model.transcribe(..., regroup='co=compression ratio+>+3.0+remove+0_da')

For CLI: --regroup "co=compression ratio+>+3.0+remove+0_da"