In this project, I attempted to explore the model Whisper-Tiny from OpenAI, which is a speech transcription model that takes in audio and provides a transcript as output. Here’s an example of transcription

Reference text: CHAPTER SIXTEEN I MIGHT HAVE TOLD YOU OF THE BEGINNING OF THIS LIAISON IN A FEW LINES BUT I WANTED YOU TO SEE EVERY STEP BY WHICH WE CAME I TO AGREE TO WHATEVER MARGUERITE WISHED
Transcription:  Chapter 16 I might have told you of the beginning of this liaison in a few lines, but I wanted you to see every step by which we came. I to agree to whatever Mark Reid wished.
Input features shape: torch.Size([80, 3000])
Audio duration: 14.53 seconds

Residual Stream Analysis

Whisper is an encoder-decoder style model. We begin by generating some statistics on the residual stream, we mainly look at activation distributions in the last layer of of the encoder across 5 examples and see that the distributions are rather uniform in nature. Due to the nature of the space of these activations (384-dimensions) and the fact that they are non-sparse, these are rather difficult to interpret. We will therefore, later in this document show how we use sparse autoencoders (SAEs) to extract more interpretable features from this space.

Now, we analyze some attention patterns from the encoder based on an example.

We use Claude to generate an explanation of these patterns and verify them

Layer 1 Attention (First Head):
- Shows a clear diagonal pattern, indicating that each position is primarily attending to itself and nearby positions.
- The diagonal is thin and sharp, suggesting a very focused local attention.
Layer 2 Attention (First Head):
- The diagonal pattern is still present but slightly broader.
- There's more diffusion of attention around the diagonal, indicating the model is considering a wider context.
Layer 3 Attention (First Head):
- The diagonal is even more diffuse, with attention spreading further from the main diagonal.
- This suggests the model is integrating information from a broader context in this layer.
Layer 4 Attention (First Head):
- The attention pattern is significantly different from the previous layers.
- There's a strong focus on the early part of the sequence (top-left corner), with some vertical striping.
- This could indicate that the model is using this layer to aggregate information from the entire sequence, with a bias towards earlier elements.

This gr...

Posts

Wikitag Contributions

Comments

Posts

Wikitag Contributions

Comments