Self-Attentional Acoustic Models

TL;DR

This paper explores applying self-attention to acoustic modeling, addressing computational challenges with novel techniques, and demonstrating that self-attentional models can be faster, interpretable, and comparable to LSTM baselines.

Contribution

The paper introduces methods to mitigate quadratic memory growth, explores effective position encoding, and emphasizes local context control in self-attentional acoustic models.

Findings

Model approaches a strong LSTM baseline in accuracy.

Self-attentional models are significantly faster to compute.

Attention heads learn linguistically meaningful divisions.

Abstract

Self-attention is a method of encoding sequences of vectors by relating these vectors to each-other based on pairwise similarities. These models have recently shown promising results for modeling discrete sequences, but they are non-trivial to apply to acoustic modeling due to computational and modeling issues. In this paper, we apply self-attention to acoustic modeling, proposing several improvements to mitigate these issues: First, self-attention memory grows quadratically in the sequence length, which we address through a downsampling technique. Second, we find that previous approaches to incorporate position information into the model are unsuitable and explore other representations and hybrid models to this end. Third, to stress the importance of local context in the acoustic signal, we propose a Gaussian biasing approach that allows explicit control over the context range. Experiments find that our model approaches a strong baseline based on LSTMs with network-in-network connections while being much faster to compute. Besides speed, we find that interpretability is a strength of self-attentional acoustic models, and demonstrate that self-attention heads learn a linguistically plausible division of labor.