Incorrect feature_size in preprocessor_config.json (should be 80)

#199
by alexg1802 - opened

Hello,

It seems the preprocessor_config.json file for openai/whisper-large-v3 currently contains:

"feature_size": 128

However, Whisper models use 80 mel filterbanks, and this causes runtime crashes in pipelines using WhisperProcessor.from_pretrained.

Expected value: 80

Can you please correct this? Thank you!

Hi, as they mentioned in the document, Whisper large v3 comes with a 128 mel spectrogram bin size in contrast to the other models.

Sign up or log in to comment