Interpreting Pretrained Speech Models for Automatic Speech Assessment of Voice Disorders

TL;DR

This paper investigates how pretrained speech models, specifically Audio Spectrogram Transformers, make predictions for voice disorder detection by analyzing their attention mechanisms, revealing that fine-tuning concentrates attention on relevant phoneme regions.

Contribution

It introduces the use of attention rollout to interpret pretrained speech models in voice disorder detection and compares different configurations of Audio Spectrogram Transformers.

Findings

Attention becomes more focused on phoneme regions after fine-tuning.

Model relevance maps reveal how models make predictions under different conditions.

Fine-tuning reduces the spread of attention across the spectrogram.

Abstract

Speech contains information that is clinically relevant to some diseases, which has the potential to be used for health assessment. Recent work shows an interest in applying deep learning algorithms, especially pretrained large speech models to the applications of Automatic Speech Assessment. One question that has not been explored is how these models output the results based on their inputs. In this work, we train and compare two configurations of Audio Spectrogram Transformer in the context of Voice Disorder Detection and apply the attention rollout method to produce model relevance maps, the computed relevance of the spectrogram regions when the model makes predictions. We use these maps to analyse how models make predictions in different conditions and to show that the spread of attention is reduced as a model is finetuned, and the model attention is concentrated on specific phoneme regions.