It appears that compatibility issues persist between whisper_timestamped v1.15.5 and openai-whisper v20240930 (TypeError: 'NoneType' object is not subscriptable)

[JHjang223]

I'm trying to use whisper-timestamped v1.15.5 along with the latest version of openai-whisper v20240930. However, after installing both versions, I encountered the following error when running the code( Additionally, i am using Python 3.9.20 and pytorch 2.1.2+cu121:

Traceback (most recent call last):
  File "/data1/whisper_test/inference.py", line 11, in test_whisper_inf
    segments = whisper.transcribe(model, file_path, language="ko", beam_size=5, best_of=5, temperature=(0.0, 0.2, 0.4, 0.6, 0.8, 1.0), vad=True, detect_disfluencies=True)
  File "/data1/whisper_test/whisper_timestamped/transcribe.py", line 304, in transcribe_timestamped
    (transcription, words) = _transcribe_timestamped_naive(model, audio,
  File "/data1/whisper_test/whisper_timestamped/transcribe.py", line 1246, in _transcribe_timestamped_naive
    attention_weights = [w[:, :, i_start-1:, :] for w in attention_weights]
  File "/data1/whisper_test/whisper_timestamped/transcribe.py", line 1246, in <listcomp>
    attention_weights = [w[:, :, i_start-1:, :] for w in attention_weights]
TypeError: 'NoneType' object is not subscriptable

Based on my investigation, it appears that this error is caused by the selective addition of a function in openai-whisper v20240930 that does not return attention weights (qk). Specifically, the scaled_dot_product_attention in torch.nn.functional was applied in the v20240930 update, and since this function does not return attention weights (qk), it seems to be causing compatibility issues with whisper-timestamped.

Although I confirmed #212, I've noticed that the disabled_sdpa() function in the latest whisper-timestamped v1.15.5 does not seem to resolve this issue effectively. Anyway, The code works perfectly with openai-whisper v20240927, where qk values are still always returned.