Segment timestamps are buggy in BatchedInferencePipeline

[kalradivyanshu]

In the recent code #856, BatchedInferencePipeline the segments are sometimes wrong for some reason.

I recorded an audio clip: https://drive.google.com/file/d/1cbDbiXi12SIsd0hIDfs61VdtgI78Fg_p/view?usp=sharing

and for this BatchedInferencePipeline gives Segment(id=1, seek=2307, start=18.71, end=23.07... even though the clip is only 3 second long.

Here is a colab recreating this issue, just upload batch.wav from the link above: https://colab.research.google.com/drive/1ie7uMFW_LJUvxGHW3KkT5iG8uUZiTVwU?usp=sharing

segments, info = batched_model.transcribe(arr, word_timestamps=True, batch_size = 1)
print(list(segments))

output:

[Segment(id=1, seek=2307, start=18.71, end=23.07, text=' Hey Michael, how are you?', tokens=[14690, 3899, 11, 703, 389, 345, 30], avg_logprob=-0.21215820871293545, compression_ratio=0.7647058823529411, no_speech_prob=0.08087158203125, words=[Word(start=18.71, end=20.11, word=' Hey', probability=0.71240234375), Word(start=20.11, end=20.11, word=' Michael,', probability=0.732421875), Word(start=22.71, end=22.71, word=' how', probability=0.98046875), Word(start=22.71, end=22.71, word=' are', probability=0.9990234375), Word(start=22.71, end=23.07, word=' you?', probability=0.99853515625)], temperature=1.0)]

segments, info = model.transcribe(arr, word_timestamps=True)
print(list(segments))

correct output:

[Segment(id=1, seek=330, start=0.9999999999999996, end=2.5, text=' Hey Michael, how are you?', tokens=[50413, 14690, 3899, 11, 703, 389, 345, 30, 50513], avg_logprob=-0.568749976158142, compression_ratio=0.7575757575757576, no_speech_prob=0.078125, words=[Word(start=0.9999999999999996, end=1.4, word=' Hey', probability=0.7177734375), Word(start=1.4, end=1.6, word=' Michael,', probability=0.68896484375), Word(start=2.04, end=2.14, word=' how', probability=0.98046875), Word(start=2.14, end=2.32, word=' are', probability=0.998046875), Word(start=2.32, end=2.5, word=' you?', probability=0.99755859375)], temperature=0.0)]

Thank you for all your work!

[kalradivyanshu]

Did some digging around, I think the words timestamp is the problem.

In these lines:
https:/SYSTRAN/faster-whisper/blob/master/faster_whisper/transcribe.py#L1794-L1814

The last word's end timestamp is used for segments end timestamp. Which points to the error being in the find_alignment function:

faster-whisper/faster_whisper/transcribe.py

Line 1820 in eb83902

def find_alignment(

This can be confirmed by the fact that if I don't send word_timestamps=True the segment timestamps are correct.

[Jiltseb]

@MahmoudAshraf97

[MahmoudAshraf97]

should be fixed in #920

[kalradivyanshu]

Hey @MahmoudAshraf97 Thank you for your quick response and fix!

I have been able to recreate it in your code, but for a specific case:
audio: https://drive.google.com/file/d/1oK0x7OF_JrfWa7ot1-bkTSP1m_tP_vEk/view?usp=sharing

colab: https://colab.research.google.com/drive/1ie7uMFW_LJUvxGHW3KkT5iG8uUZiTVwU?usp=sharing

In this, it works if I use batched_model.transcribe's VAD, but if I pass in VAD segments specifically: [{'start': 0.0, 'end': 30.0, 'segments': [(0.832, 7.024000000000001)] It breaks:

[Segment(id=1, seek=2998, start=0.0, end=29.98, text=" Good morning, Hank, it's Tuesday. You know how they're those videos that are like so-and-so answers the web's most searched questions about them, and", tokens=[4599, 3329, 11, 24386, 11, 340, 338, 3431, 13, 921, 760, 703, 484, 821, 883, 5861, 326, 389, 588, 523, 12, 392, 12, 568, 7429, 262, 3992, 338, 749, 16499, 2683, 546, 606, 11, 290], avg_logprob=-0.22146267195542654, compression_ratio=1.2295081967213115, no_speech_prob=0.03997802734375, words=[Word(start=0.0, end=0.1, word=' Good', probability=0.87646484375), Word(start=0.1, end=0.3, word=' morning,', probability=0.91845703125), Word(start=0.34, end=0.46, word=' Hank,', probability=0.409423828125), Word(start=0.46, end=0.62, word=" it's", probability=0.983154296875), Word(start=0.62, end=0.74, word=' Tuesday.', probability=0.99755859375), Word(start=1.02, end=1.02, word=' You', probability=0.97119140625), Word(start=1.02, end=1.1, word=' know', probability=0.99658203125), Word(start=1.1, end=1.22, word=' how', probability=0.87548828125), Word(start=1.22, end=1.4, word=" they're", probability=0.70458984375), Word(start=1.4, end=1.54, word=' those', probability=0.943359375), Word(start=1.54, end=1.86, word=' videos', probability=0.99560546875), Word(start=1.86, end=2.1, word=' that', probability=0.974609375), Word(start=2.1, end=2.22, word=' are', probability=0.9951171875), Word(start=2.22, end=2.38, word=' like', probability=0.66796875), Word(start=2.38, end=2.68, word=' so', probability=0.7060546875), Word(start=2.68, end=2.82, word='-and', probability=0.852294921875), Word(start=2.82, end=29.98, word='-so', probability=0.996337890625), Word(start=29.98, end=29.98, word=' answers', probability=0.98046875), Word(start=29.98, end=29.98, word=' the', probability=0.962890625), Word(start=29.98, end=29.98, word=" web's", probability=0.869140625), Word(start=29.98, end=29.98, word=' most', probability=0.9443359375), Word(start=29.98, end=29.98, word=' searched', probability=0.9453125), Word(start=29.98, end=29.98, word=' questions', probability=0.99365234375), Word(start=29.98, end=29.98, word=' about', probability=0.9990234375), Word(start=29.98, end=29.98, word=' them,', probability=0.9970703125), Word(start=29.98, end=29.98, word=' and', probability=0.9873046875)], temperature=1.0)]

However if I pass the same segments into normal transcribe, using clip_timestamps:

segments, info = model.transcribe(arr, word_timestamps=True, clip_timestamps = [0.832, 7.024000000000001], vad_filter=False)

That works correctly:

[Segment(id=1, seek=702, start=0.83, end=5.93, text=" You know how they're those videos that are like so-and-so answers the web's most searched questions about them and", tokens=[50363, 921, 760, 703, 484, 821, 883, 5861, 326, 389, 588, 523, 12, 392, 12, 568, 7429, 262, 3992, 338, 749, 16499, 2683, 546, 606, 290, 50619], avg_logprob=-0.22739954824958528, compression_ratio=1.1875, no_speech_prob=0.02056884765625, words=[Word(start=0.83, end=0.99, word=' You', probability=0.37646484375), Word(start=0.99, end=1.11, word=' know', probability=0.99267578125), Word(start=1.11, end=1.23, word=' how', probability=0.81787109375), Word(start=1.23, end=1.37, word=" they're", probability=0.66845703125), Word(start=1.37, end=1.55, word=' those', probability=0.8408203125), Word(start=1.55, end=1.87, word=' videos', probability=0.990234375), Word(start=1.87, end=2.11, word=' that', probability=0.96337890625), Word(start=2.11, end=2.21, word=' are', probability=0.9931640625), Word(start=2.21, end=2.37, word=' like', probability=0.8662109375), Word(start=2.37, end=2.69, word=' so', probability=0.7451171875), Word(start=2.69, end=2.83, word='-and', probability=0.7646484375), Word(start=2.83, end=2.99, word='-so', probability=0.99755859375), Word(start=2.99, end=3.35, word=' answers', probability=0.96826171875), Word(start=3.35, end=3.67, word=' the', probability=0.95703125), Word(start=3.67, end=4.07, word=" web's", probability=0.839111328125), Word(start=4.07, end=4.25, word=' most', probability=0.970703125), Word(start=4.25, end=4.51, word=' searched', probability=0.943359375), Word(start=4.51, end=4.87, word=' questions', probability=0.99462890625), Word(start=4.87, end=5.47, word=' about', probability=0.9970703125), Word(start=5.47, end=5.77, word=' them', probability=0.99755859375), Word(start=5.77, end=5.93, word=' and', probability=0.55078125)], temperature=0.0)]

I think its some edge condition :/

[MahmoudAshraf97]

because the input is not the same
the end of a vad segment should never be more than the end of its last subsegment, in your case the end is 30.0 while the subsegment end is 7.024
this is equivalent to
clip_timestamps = [0.0, 30.0]

[kalradivyanshu]

Oh ok, my bad, I fixed that, and it is working now!