Improve language detection when using clip_timestamps

[ben91lin]

Using clip_timestamps to improve the initial seek for language detection, avoiding incorrect detection at the start of the audio file.

[trungkienbkhn]

Should replace with:

seek = int(start_timestamp * self.frames_per_second)

[trungkienbkhn]

@ben91lin , hello. Tks for your improvement. But I found that an error occurs if clip_timestamp[0] > audio duration:

    language = max(
ValueError: max() arg is an empty sequence

Could you fix this ?

[ben91lin]

# If seek is beyond all frames, set it to the last segment.
if seek >= features.shape[-1]:
    seek = content_frames

If audio_length is 80s and start_timestamp is 67s, it will clip the last 1.3 seconds for detection.
If start_timestamp greater or equal 80s, force use the last nb_max_frames for detection.

[trungkienbkhn]

# If seek is beyond all frames, set it to the last segment.
if seek >= features.shape[-1]:
    seek = content_frames

I think it's wrong logic. In this case, if seek = content_frame, then fw will detect language from content_frames to features.shape[-1].
=> These are padded values (meaningless), and can lead to incorrect language detection.
Should set seek = 0 if seek >= content_frames

[ben91lin]

I think you are right, I omitted the padding from FeatureExtractor. Setting seek = 0 if seek >= content_frame is the correct approach.