feat: return avg_logprobs from faster_whisper

[werkamsus]

faster-whisper can output avg_logsprobs - this is a patch to pass them to whisperx.transcribe.
kept changes minimal & tested everything.

note: I couldn't figure out how to pass the avg_logsprobs to alignment / diarization & built a workaround for myself, you can call it after alignment / diarization if you save the initial result + the last result:

def match_segments(first_dict, second_dict):
    aligned_segments = []

    for first_segment in first_dict['segments']:
        aligned_segment = None
        closest_difference = float('inf')

        for second_segment in second_dict['segments']:
            # Compute the difference in start and end times
            start_difference = abs(first_segment['start'] - second_segment['start'])
            end_difference = abs(first_segment['end'] - second_segment['end'])
            total_difference = start_difference + end_difference

            # Check if this segment is the closest match
            if total_difference < closest_difference:
                closest_difference = total_difference
                aligned_segment = second_segment.copy()

        # Add avg_logprob to the closest matching segment
        if aligned_segment:
            aligned_segment['avg_logprob'] = first_segment['avg_logprob']
            aligned_segments.append(aligned_segment)

    return {'segments': aligned_segments}


matched_segments = match_segments(result, result_align)

[pprobst]

This is very relevant!

[Tejaswgupta]

@m-bain would you be merging this one in?

[lfcnassif]

+1

Thanks for this great project! Can this PR be used in production? We have a hard requirement to return to our users the transcription average confidence score.

[lfcnassif]

For those interested, I copied this commit over current last main commit into this fork and branch:
https://github.com/sepinf-inc/whisperX/tree/confidence_score