Visualize attention maps on input spectrograms

[mmerler]

Does anyone know how to visualize the encoder attention maps with respect to the input spectrograms?
I'm interested in understanding which portions of the spectrogram a whisper-base fine-tuned model is focusing on when making a prediction.
I can extract the attention maps in the forward pass, each is 1500x1500, but I don't know how to map them back to the input spectrogram.

Any ideas?

[mmerler]

basically the equivalent of Grad-Cam for audio with whisper?

[Coder1010ayush]

Any updates on this?