Whisper - list index out of range with word level timestamps

[maxkvbn]

System Info

• transformers version: 4.42.2
• Platform: Windows-10-10.0.22621-SP0
• Python version: 3.10.14
• Huggingface_hub version: 0.23.4
• Safetensors version: 0.4.2
• Accelerate version: 0.31.0
• Accelerate config: not found
• PyTorch version (GPU?): 2.3.1 (True)
• Tensorflow version (GPU?): not installed (NA)
• Flax version (CPU?/GPU?/TPU?): not installed (NA)
• Jax version: not installed
• JaxLib version: not installed
• Using distributed or parallel set-up in script?: No
• Using GPU in script?: Yes
• GPU type: NVIDIA GeForce RTX 4070 Laptop GPU

Who can help?

No response

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
• My own task or dataset (give details below)

Reproduction

1. Load 'whisper-large-v3' AutoModelForSpeechSeq2Seq model and send it to GPU
2. Setup model config with return_timestamps="word", amongst other settings.
3. Run audio through the pipe, which results in error

---------------------------------------------------------------------------
IndexError                                Traceback (most recent call last)
Cell In[8], [line 1](vscode-notebook-cell:?execution_count=8&line=1)
----> [1](vscode-notebook-cell:?execution_count=8&line=1) asr_out = transcribing_pipe(
      [2](vscode-notebook-cell:?execution_count=8&line=2)     '../SampleData/Saba_interview_short.wav',
      [3](vscode-notebook-cell:?execution_count=8&line=3)     return_timestamps="word",
      [4](vscode-notebook-cell:?execution_count=8&line=4)     generate_kwargs={"language": "danish"}
      [5](vscode-notebook-cell:?execution_count=8&line=5)     )
      [7](vscode-notebook-cell:?execution_count=8&line=7) asr_out

File c:\Users\User\miniconda3\envs\vva\lib\site-packages\transformers\pipelines\automatic_speech_recognition.py:284, in AutomaticSpeechRecognitionPipeline.__call__(self, inputs, **kwargs)
    [221](file:///C:/Users/User/miniconda3/envs/vva/lib/site-packages/transformers/pipelines/automatic_speech_recognition.py:221) def __call__(
    [222](file:///C:/Users/User/miniconda3/envs/vva/lib/site-packages/transformers/pipelines/automatic_speech_recognition.py:222)     self,
    [223](file:///C:/Users/User/miniconda3/envs/vva/lib/site-packages/transformers/pipelines/automatic_speech_recognition.py:223)     inputs: Union[np.ndarray, bytes, str],
    [224](file:///C:/Users/User/miniconda3/envs/vva/lib/site-packages/transformers/pipelines/automatic_speech_recognition.py:224)     **kwargs,
    [225](file:///C:/Users/User/miniconda3/envs/vva/lib/site-packages/transformers/pipelines/automatic_speech_recognition.py:225) ):
    [226](file:///C:/Users/User/miniconda3/envs/vva/lib/site-packages/transformers/pipelines/automatic_speech_recognition.py:226)     """
    [227](file:///C:/Users/User/miniconda3/envs/vva/lib/site-packages/transformers/pipelines/automatic_speech_recognition.py:227)     Transcribe the audio sequence(s) given as inputs to text. See the [`AutomaticSpeechRecognitionPipeline`]
    [228](file:///C:/Users/User/miniconda3/envs/vva/lib/site-packages/transformers/pipelines/automatic_speech_recognition.py:228)     documentation for more information.
   (...)
    [282](file:///C:/Users/User/miniconda3/envs/vva/lib/site-packages/transformers/pipelines/automatic_speech_recognition.py:282)                 `"".join(chunk["text"] for chunk in output["chunks"])`.
    [283](file:///C:/Users/User/miniconda3/envs/vva/lib/site-packages/transformers/pipelines/automatic_speech_recognition.py:283)     """
--> [284](file:///C:/Users/User/miniconda3/envs/vva/lib/site-packages/transformers/pipelines/automatic_speech_recognition.py:284)     return super().__call__(inputs, **kwargs)

File c:\Users\User\miniconda3\envs\vva\lib\site-packages\transformers\pipelines\base.py:1246, in Pipeline.__call__(self, inputs, num_workers, batch_size, *args, **kwargs)
   [1244](file:///C:/Users/User/miniconda3/envs/vva/lib/site-packages/transformers/pipelines/base.py:1244)     return self.iterate(inputs, preprocess_params, forward_params, postprocess_params)
   [1245](file:///C:/Users/User/miniconda3/envs/vva/lib/site-packages/transformers/pipelines/base.py:1245) elif self.framework == "pt" and isinstance(self, ChunkPipeline):
-> [1246](file:///C:/Users/User/miniconda3/envs/vva/lib/site-packages/transformers/pipelines/base.py:1246)     return next(
   [1247](file:///C:/Users/User/miniconda3/envs/vva/lib/site-packages/transformers/pipelines/base.py:1247)         iter(
   [1248](file:///C:/Users/User/miniconda3/envs/vva/lib/site-packages/transformers/pipelines/base.py:1248)             self.get_iterator(
   [1249](file:///C:/Users/User/miniconda3/envs/vva/lib/site-packages/transformers/pipelines/base.py:1249)                 [inputs], num_workers, batch_size, preprocess_params, forward_params, postprocess_params
   [1250](file:///C:/Users/User/miniconda3/envs/vva/lib/site-packages/transformers/pipelines/base.py:1250)             )
   [1251](file:///C:/Users/User/miniconda3/envs/vva/lib/site-packages/transformers/pipelines/base.py:1251)         )
   [1252](file:///C:/Users/User/miniconda3/envs/vva/lib/site-packages/transformers/pipelines/base.py:1252)     )
   [1253](file:///C:/Users/User/miniconda3/envs/vva/lib/site-packages/transformers/pipelines/base.py:1253) else:
   [1254](file:///C:/Users/User/miniconda3/envs/vva/lib/site-packages/transformers/pipelines/base.py:1254)     return self.run_single(inputs, preprocess_params, forward_params, postprocess_params)

File c:\Users\User\miniconda3\envs\vva\lib\site-packages\transformers\pipelines\pt_utils.py:125, in PipelineIterator.__next__(self)
    [123](file:///C:/Users/User/miniconda3/envs/vva/lib/site-packages/transformers/pipelines/pt_utils.py:123) # We're out of items within a batch
    [124](file:///C:/Users/User/miniconda3/envs/vva/lib/site-packages/transformers/pipelines/pt_utils.py:124) item = next(self.iterator)
--> [125](file:///C:/Users/User/miniconda3/envs/vva/lib/site-packages/transformers/pipelines/pt_utils.py:125) processed = self.infer(item, **self.params)
    [126](file:///C:/Users/User/miniconda3/envs/vva/lib/site-packages/transformers/pipelines/pt_utils.py:126) # We now have a batch of "inferred things".
    [127](file:///C:/Users/User/miniconda3/envs/vva/lib/site-packages/transformers/pipelines/pt_utils.py:127) if self.loader_batch_size is not None:
    [128](file:///C:/Users/User/miniconda3/envs/vva/lib/site-packages/transformers/pipelines/pt_utils.py:128)     # Try to infer the size of the batch

File c:\Users\User\miniconda3\envs\vva\lib\site-packages\transformers\pipelines\automatic_speech_recognition.py:587, in AutomaticSpeechRecognitionPipeline.postprocess(self, model_outputs, decoder_kwargs, return_timestamps, return_language)
    [584](file:///C:/Users/User/miniconda3/envs/vva/lib/site-packages/transformers/pipelines/automatic_speech_recognition.py:584)             stride_right /= sampling_rate
    [585](file:///C:/Users/User/miniconda3/envs/vva/lib/site-packages/transformers/pipelines/automatic_speech_recognition.py:585)             output["stride"] = chunk_len, stride_left, stride_right
--> [587](file:///C:/Users/User/miniconda3/envs/vva/lib/site-packages/transformers/pipelines/automatic_speech_recognition.py:587)     text, optional = self.tokenizer._decode_asr(
    [588](file:///C:/Users/User/miniconda3/envs/vva/lib/site-packages/transformers/pipelines/automatic_speech_recognition.py:588)         model_outputs,
    [589](file:///C:/Users/User/miniconda3/envs/vva/lib/site-packages/transformers/pipelines/automatic_speech_recognition.py:589)         return_timestamps=return_timestamps,
    [590](file:///C:/Users/User/miniconda3/envs/vva/lib/site-packages/transformers/pipelines/automatic_speech_recognition.py:590)         return_language=return_language,
    [591](file:///C:/Users/User/miniconda3/envs/vva/lib/site-packages/transformers/pipelines/automatic_speech_recognition.py:591)         time_precision=time_precision,
    [592](file:///C:/Users/User/miniconda3/envs/vva/lib/site-packages/transformers/pipelines/automatic_speech_recognition.py:592)     )
    [593](file:///C:/Users/User/miniconda3/envs/vva/lib/site-packages/transformers/pipelines/automatic_speech_recognition.py:593) else:
    [594](file:///C:/Users/User/miniconda3/envs/vva/lib/site-packages/transformers/pipelines/automatic_speech_recognition.py:594)     items = np.concatenate(final_items, axis=1)

File c:\Users\User\miniconda3\envs\vva\lib\site-packages\transformers\models\whisper\tokenization_whisper.py:832, in WhisperTokenizer._decode_asr(self, model_outputs, return_timestamps, return_language, time_precision)
    [831](file:///C:/Users/User/miniconda3/envs/vva/lib/site-packages/transformers/models/whisper/tokenization_whisper.py:831) def _decode_asr(self, model_outputs, *, return_timestamps, return_language, time_precision):
--> [832](file:///C:/Users/User/miniconda3/envs/vva/lib/site-packages/transformers/models/whisper/tokenization_whisper.py:832)     return _decode_asr(
    [833](file:///C:/Users/User/miniconda3/envs/vva/lib/site-packages/transformers/models/whisper/tokenization_whisper.py:833)         self,
    [834](file:///C:/Users/User/miniconda3/envs/vva/lib/site-packages/transformers/models/whisper/tokenization_whisper.py:834)         model_outputs,
    [835](file:///C:/Users/User/miniconda3/envs/vva/lib/site-packages/transformers/models/whisper/tokenization_whisper.py:835)         return_timestamps=return_timestamps,
    [836](file:///C:/Users/User/miniconda3/envs/vva/lib/site-packages/transformers/models/whisper/tokenization_whisper.py:836)         return_language=return_language,
    [837](file:///C:/Users/User/miniconda3/envs/vva/lib/site-packages/transformers/models/whisper/tokenization_whisper.py:837)         time_precision=time_precision,
    [838](file:///C:/Users/User/miniconda3/envs/vva/lib/site-packages/transformers/models/whisper/tokenization_whisper.py:838)     )

File c:\Users\User\miniconda3\envs\vva\lib\site-packages\transformers\models\whisper\tokenization_whisper.py:1032, in _decode_asr(tokenizer, model_outputs, return_timestamps, return_language, time_precision)
   [1030](file:///C:/Users/User/miniconda3/envs/vva/lib/site-packages/transformers/models/whisper/tokenization_whisper.py:1030) current_tokens.append(token)
   [1031](file:///C:/Users/User/miniconda3/envs/vva/lib/site-packages/transformers/models/whisper/tokenization_whisper.py:1031) if return_timestamps == "word":
-> [1032](file:///C:/Users/User/miniconda3/envs/vva/lib/site-packages/transformers/models/whisper/tokenization_whisper.py:1032)     start_time = round(token_timestamps[i] + time_offset, 2)
   [1033](file:///C:/Users/User/miniconda3/envs/vva/lib/site-packages/transformers/models/whisper/tokenization_whisper.py:1033)     if i + 1 < len(token_timestamps):
   [1034](file:///C:/Users/User/miniconda3/envs/vva/lib/site-packages/transformers/models/whisper/tokenization_whisper.py:1034)         end_time = round(token_timestamps[i + 1] + time_offset, 2)

IndexError: list index out of range

I've uploaded the audio that I am trying to process here

I have a discussion on the whisper's hub page, where I have implemented a fix that seems to work. here

Colab link here

Expected behavior

1. Load 'whisper-large-v3' AutoModelForSpeechSeq2Seq model and send it to GPU
2. Setup model config with return_timestamps="word", amongst other settings.
3. Run audio through the pipe, which returns transcription together with the word level timestamps segment.

[amyeroberts]

cc @sanchit-gandhi @kamilakesbi