Update logic to get segment from features before encoding

[trungkienbkhn]

Quick fix for #646

I tested with a poor quality audio file (192s) and received:

• ctranslate2==3.24.0 and faster-whisper==0.10.1, the execution time is 12.9s
• ctranslate2==3.24.0 and faster-whisper==1.0.1, the execution time is 28.9s

Below is my testing code:

model = WhisperModel('large-v3', device='cuda')
segments, info = model.transcribe(jfk_path, word_timestamps=True)

After investigating, I found that the logic changes below increased the latency:
Old logic:

segment = features[:, seek : seek + self.feature_extractor.nb_max_frames]
segment_size = min(
    self.feature_extractor.nb_max_frames, content_frames - seek
)

New update logic in #646:

segment_size = min(
    self.feature_extractor.nb_max_frames,
    content_frames - seek,
    seek_clip_end - seek,
)
segment = features[:, seek : seek + segment_size]

Cutting the segment like this changed the encode_output and it only happens at the last loop because then segment_size < self.feature_extractor.nb_max_frames.
So the decode_result also changed and it's quality is reduced in my tests.
Specifically, I got the following error in my log:

Compression ratio threshold is not met with temperature 0.0 (3.9903846153846154 > 2.4)
Compression ratio threshold is not met with temperature 0.2 (3.9903846153846154 > 2.4)
Compression ratio threshold is not met with temperature 0.4 (3.9903846153846154 > 2.4)
Compression ratio threshold is not met with temperature 0.6 (3.7184466019417477 > 2.4)

=> The quality of decode_result is poor and repetitive so it takes more iterations to decode and export final result, leading to increased transcription time.
For more info, you can check this logic.

[thomasmol]

I tried this and fixes a latency/speed bug for me mentioned in #703
Thanks!

[trungkienbkhn]

@kale4eat, tks for your recommend in #712. I updated my code logic with pad_or_trim function and it works as expected.

[doublex]

@trungkienbkhn
Thanks for your efforts!