Concentrating Harder for Faster Audio Transformer

Abstract

Attention-based models have become tremendously successful in the last couple of years for tasks such as Acoustic Scene Classification, Event Classification, or Speaker Identification. They operate on a set of tokens extracted from audio features and scale quadratically in sequence length due to their pairwise operations in multi-head attention.In this paper, we propose to parameterize a Dirichlet prior with a Transformer model, to jointly estimate label and token probabilities. A token bottleneck, provided by a Categorical-Dirichlet pair, forces the model to concentrate on a subset of tokens. This allows for improved interpretability and higher audio throughput during inference. We compare two different methods for token sampling – full knowledge of all tokens and partial token sampling for reduced complexity.We evaluate and interpret typical audio datasets such as the Environmental Sound Classification (ESC-50) dataset, the TAU Urban Acoustic Scenes 2020 (TAU20) dataset, and the Speech Command version 2 (SCv2) dataset. The results show that the token budget can be reduced without significant performance loss, especially for acoustic scene classification. We show that for the ESC-50 and SCv2 datasets, the token relevance can be well approximated with partial token view. Finally, we show that a significant increase in throughput can be achieved with our proposed methods.