Attention-Based Models for Speech Recognition

TL;DR

This paper enhances attention-based models for speech recognition by introducing location-aware mechanisms, improving robustness to long inputs and reducing phoneme error rates significantly.

Contribution

It proposes a novel location-aware attention mechanism for speech recognition, addressing length sensitivity issues in existing models.

Findings

Achieved 18% PER on TIMIT with initial adaptation

Enhanced model robustness to longer utterances

Reduced PER to 17.6% with improved attention mechanism

Abstract

Recurrent sequence generators conditioned on input data through an attention mechanism have recently shown very good performance on a range of tasks in- cluding machine translation, handwriting synthesis and image caption gen- eration. We extend the attention-mechanism with features needed for speech recognition. We show that while an adaptation of the model used for machine translation in reaches a competitive 18.7% phoneme error rate (PER) on the TIMIT phoneme recognition task, it can only be applied to utterances which are roughly as long as the ones it was trained on. We offer a qualitative explanation of this failure and propose a novel and generic method of adding location-awareness to the attention mechanism to alleviate this issue. The new method yields a model that is robust to long inputs and achieves 18% PER in single utterances and 20% in 10-times longer (repeated) utterances. Finally, we propose a change to the at- tention mechanism that prevents it from concentrating too much on single frames, which further reduces PER to 17.6% level.