#33512 handle last element out of range error

[itazap]

#33552
fix to handle out of range error

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[ArthurZucker]

Thanks! let's add the issues' reproducer as a small test!

[itazap]

@ArthurZucker original issue used stable_whisper and can't reproduce the problem if loading from WhisperTokenizer, not sure if we should add the lib dependency for a stable_whisper test?

[ylacombe]

Hey there, would it be possible to have a snippet to reproduce the issue ? It might actually be an issue with Whisper modeling code rather than on the tokenizer side.

cc @eustlb

[itazap]

@ylacombe @eustlb yes snippet and audio file is here: #33552

or you can run the below:

EDIT: I think I've pinpointed the issue. The first sequence does not end with a token in all_special_ids. Is this a Whisper requirement?

model_outputs = [{'tokens': np.array([[50258, 50259, 50359, 50364,  1468,   380,   976,   385,   300,   286,
           312,  2633,   300,   472,   551,   300,   286,  2378,   380,  2762,
           294,  5680,    13, 20180,   281,  1446,   300,  6944,  4931,   281,
           841,   512,   777,  5503,    13,   583,  1968,   420,   406,   286,
           603,  3270,   689,   286,  2117,    13, 10865,   286,   478,   445,
          1382,   281,  5268,    13,  2432, 40128,   521,   264,   777,  1393,
          3082,   286,   478,   257,  3429,   586,    13,   316, 12232,   295,
           445,   281,  1621,   926,   309,    13, 34695,   271,    13,  5303,
           257,   274,  3019,   281,  1855,   293,  1087,   484,    13,  4055,
           266, 21065,  4570,   286,  4244,   466,    13,   759,   436,   536,
           385,  6588,    13, 30308,   264, 21065,  1626,  9019,   466,    13,
          8503,   286,   478,  2633,   760,   420,  2633,   264,   558,  2372,
            13,   286,   841,   264,   596,   346,    13,   583,   406,  1547,
           281]]), 'token_timestamps': np.array([[ 0.0000,  0.0000,  0.0000,  0.0000,  0.0000,  0.3200,  0.3400,  0.5000,
          0.6600,  0.8600,  1.0000,  1.2600,  1.5000,  1.7600,  2.0400,  2.4000,
          2.5800,  2.8000,  2.9200,  3.0600,  3.2200,  3.5800,  4.4000,  5.5000,
          5.7800,  5.9200,  6.0800,  6.3000,  6.6400,  6.8000,  6.9800,  7.1600,
          7.4000,  7.6200,  8.1200,  8.2600,  8.4600,  8.7200,  9.0600,  9.1600,
          9.5400,  9.5800,  9.6600,  9.9200, 10.3000, 10.7200, 10.8400, 11.0000,
         11.1400, 11.1600, 11.4400, 11.6400, 11.7800, 12.2800, 12.5400, 12.7400,
         12.8600, 13.0200, 13.1800, 13.4000, 13.6800, 13.9000, 14.1000, 14.2400,
         14.4200, 14.5600, 14.8200, 14.8800, 15.1000, 15.3200, 15.4600, 15.6400,
         15.8200, 15.9800, 16.2800, 16.5000, 16.5800, 16.8000, 17.1400, 17.1400,
         17.4600, 17.6000, 17.7000, 17.7400, 17.8800, 18.1200, 18.3600, 18.5400,
         18.7800, 19.0200, 19.2400, 19.3800, 19.5600, 19.8600, 20.0600, 20.2800,
         20.5800, 20.8000, 20.9400, 21.1600, 21.3200, 21.5600, 21.7400, 21.8600,
         22.1000, 22.3200, 22.5400, 22.7200, 23.1800, 23.5000, 23.7400, 23.8800,
         24.0600, 24.0600, 24.3000, 24.5400, 24.7200, 24.9800, 25.2000, 25.3600,
         25.6200, 25.6600, 25.8000, 26.0600, 26.2600, 26.3400, 26.4800, 26.5200,
         26.7200, 26.8600, 27.0800]]), 'stride': (30.0, 0.0, 5.0)}, {'tokens': np.array([[50258, 50259, 50359, 50363,  1449,   466,   498,   436,   536,   385,
          6588,  3974,   257,  2307,  1626,  9019,   466,  1968,   286,   478,
          2633,  6385,   286,   478,  2633,   264,   558,  2372,   286,   841,
           264,  4588,   457,   406,  1547,   281,   652,   385,   605,   493,
           411,   257, 22209,    13,   286,   478,  2633,   760,  3275,    13,
           492,   434,  2633, 50257, 50257, 50257, 50257, 50257, 50257, 50257,
         50257, 50257, 50257, 50257, 50257, 50257, 50257, 50257, 50257, 50257,
         50257, 50257, 50257, 50257, 50257, 50257, 50257, 50257, 50257, 50257,
         50257, 50257, 50257, 50257, 50257, 50257, 50257, 50257, 50257, 50257,
         50257, 50257, 50257, 50257, 50257, 50257, 50257, 50257, 50257, 50257,
         50257, 50257, 50257, 50257, 50257, 50257, 50257, 50257, 50257, 50257,
         50257, 50257, 50257, 50257, 50257, 50257, 50257, 50257, 50257, 50257,
         50257, 50257, 50257, 50257, 50257, 50257, 50257, 50257, 50257, 50257,
         50257]]), 'token_timestamps': np.array([[0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.2600, 0.6800, 0.9800, 1.2400,
         1.3000, 1.5600, 1.8400, 2.1200, 2.3600, 2.5600, 2.7400, 3.1600, 3.6000,
         3.8800, 4.1000, 4.1200, 4.3200, 4.5800, 4.7600, 4.8400, 4.9000, 5.2000,
         5.3600, 5.6200, 5.8200, 6.0200, 6.2600, 6.4800, 6.7400, 6.8600, 7.0800,
         7.3200, 7.4200, 7.6600, 7.8000, 8.0000, 8.2200, 8.4200, 8.5200, 8.5200,
         8.6800, 8.8200, 9.0000, 9.4800, 9.7000, 9.7600, 9.9000, 9.9600, 9.9800,
         9.9800, 9.9800, 9.9800, 9.9800, 9.9800, 9.9800, 9.9800, 9.9800, 9.9800,
         9.9800, 9.9800, 9.9800, 9.9800, 9.9800, 9.9800, 9.9800, 9.9800, 9.9800,
         9.9800, 9.9800, 9.9800, 9.9800, 9.9800, 9.9800, 9.9800, 9.9800, 9.9800,
         9.9800, 9.9800, 9.9800, 9.9800, 9.9800, 9.9800, 9.9800, 9.9800, 9.9800,
         9.9800, 9.9800, 9.9800, 9.9800, 9.9800, 9.9800, 9.9800, 9.9800, 9.9800,
         9.9800, 9.9800, 9.9800, 9.9800, 9.9800, 9.9800, 9.9800, 9.9800, 9.9800,
         9.9800, 9.9800, 9.9800, 9.9800, 9.9800, 9.9800, 9.9800, 9.9800, 9.9800,
         9.9800, 9.9800, 9.9800, 9.9800, 9.9800, 9.9800, 9.9800, 9.9800, 9.9800,
         9.9800, 9.9800, 9.9800, 9.9800, 9.9800]]), 'stride': (10.0075, 5.0, 0.0)}]
        # fmt: on

        tokenizer = WhisperTokenizer.from_pretrained("onnx-community/whisper-tiny.en_timestamped")
        result = tokenizer._decode_asr(
            model_outputs, return_timestamps="word", return_language=False, time_precision=0.02
        )

[itazap]

@ylacombe the tokenization code is quite complex here and I'm not familiar with the whisper model much, if you can please advise on what could be wrong in the model_ouputs or what to test , would be grateful! 😊

[ValentinKovalev]

FYI: I have tested whisper and reproduced this issue on the main branch and compared it with the current implementation. It now works correctly.

[felipehertzer]

I have tested this changes and it solve my problem with the issue #33552

[ylacombe]

Hey @itazap, sorry for the late response! Whisper's modeling code is expected to go through a number of modeling changes.

There's one particular change which deals with EOS tokens that were wrongly removed from doing short-form (i.e audio<=30s) Whisper transcription - see #33917 that'll likely be supersede by #34135.

In particular, we also have a pending question from @eustlb: should we keep or remove these special tokens ?

That said, I'm still struggling to understand why the issue you're trying to fix never appeared in our own usage of Whisper, but only in stable-whisper.

I'm also wondering if the issue also appears when doing long-form generation (i.e when audio > 30s) , which doesn't add EOS token at the end (if I remember correctly).

I believe that we should verify a few things, before actually merging this PR:

1. try to reproduce the issue with transformers-only code, to facilitate understanding of the issue. If we can't reproduce it, then we'll have to find out why it happens only in stable-whisper
2. check if the issue still occurs with Fix Whisper shortform EOS #33917 or [Whisper] 🚨 Fix whisper decoding 🚨 #34135

Depending on the answers to these questions, this PR might not be needed !

[ylacombe]

Up on this!

@eustlb, @itazap found that the first sequence does not end with a token in all_special_ids. #34537 might have solved this. Could one of you verify if it solved the issue?

If it did, let's maybe integrate the PR's test just in case.
If it didn't, it could be good that you review this PR @eustlb, as you're becoming our Whisper expert 🏎️.

[eustlb]

Finally taking a stab at this now that all the other PRs have been merged 🏗️

[itazap]

@eustlb following up after #33552 as it seems to be the common fix, let me know if it's okay to merge! 🚀

[maxkvbn]

I don't know if its just on my end, but having

return_timestamps=True

will result in the texts being empty.

Example code:

model = AutoModelForSpeechSeq2Seq.from_pretrained(
    transcribing_model_id, torch_dtype=torch_dtype, low_cpu_mem_usage=True, use_safetensors=True
)
model.to(device)
processor = AutoProcessor.from_pretrained(transcribing_model_id)

transcribing_pipe = pipeline(
    "automatic-speech-recognition",
    model=model,
    tokenizer=processor.tokenizer,
    feature_extractor=processor.feature_extractor,
    chunk_length_s=29,
    batch_size=1,
    return_timestamps=True,
    torch_dtype=torch_dtype,
    device=device,
    generate_kwargs={
                "max_new_tokens": 128,
                "condition_on_prev_tokens": True,
                "temperature": 0.2
            }
)

asr_out = transcribing_pipe(
    'example.wav',
    return_timestamps=True,
    generate_kwargs={'task': 'transcribe', 'language': 'danish'}
    )

asr_out

{'text': '',
 'chunks': [{'timestamp': (0.0, 3.6), 'text': ''},
  {'timestamp': (4.5, 6.26), 'text': ''},
  {'timestamp': (7.14, 7.54), 'text': ''},
  {'timestamp': (8.68, 6.16), 'text': ''},
  {'timestamp': (21.25, 23.41), 'text': ''},
  {'timestamp': (23.41, 25.27), 'text': ''},
  {'timestamp': (26.73, 27.01), 'text': ''},
  {'timestamp': (27.23, 28.83), 'text': ''},
  {'timestamp': (29.21, 30.57), 'text': ''},
  {'timestamp': (30.63, 31.03), 'text': ''}]}

Do you experience the same thing? @itazap

[eustlb]

@maxkvbn can you try with #36632 please ? Or provide example.wav audio if it is still failing.

[eustlb]

Hey @itazap, thank you so much for taking care of this 🤗🙏

I finally took a stab at it and solved in #36632, the approach is a bit different (see details in the PR)

[maxkvbn]

Hey @eustlb. Thanks for taking a look at this. 🙌 It seems with #36632, the TypeError issue was fixed for me. I am still getting a index out of bounds when using num_beams, but that might be a separate issue.

[eustlb]

@maxkvbn would you mind raising an issue for num_beams error please ? It's indeed a separate issue but I'd love so solve it while I'm at it 🤗

[maxkvbn]

@eustlb I see that an issue was already raised for the num_beams: #36093, so I wont create a new one. :)