Interspeech 2025 tutorial

Slides

• Grzegorz Chrupała - Introduction

• Martijn Bentum - Representational Analysis techniques

• Hosein Mohebbi - Context Mixing techniques

• Gaofei Shen - Feature Attribution techniques

• Marianne de Heer Kloots - Key takeaways & Future outlook

• Willem Zuidema - Discussion

Notebooks

We provide four notebooks, exemplifying the interpretability techniques we cover in the tutorial. The notebooks are designed to work on Google Colab, open them through the links below:

Part I: Representation Understanding

• Probing exercise: Does Wav2Vec2 encode vocal tract movements?

• Representation space comparisons: Locating acoustic and articulatory information in Wav2Vec2 with CKA

Part II: Feature Importance Scoring

• Context Mixing: Quantifying Context-Mixing in Speech Transformers

• Feature attribution: Explaining Speech Classification Models with Feature Attribution