Attention-Based Models for Speech Recognition

TLDR

Attention‑based sequence generators conditioned on input data have recently achieved strong performance on tasks such as machine translation, handwriting synthesis, and image captioning. The authors extend the attention mechanism with speech‑recognition‑specific features to enable end‑to‑end phoneme recognition. They introduce a location‑aware attention variant and a frame‑distribution regularization that prevent the model from over‑focusing on single frames. The resulting model attains 18 % phoneme error rate on single TIMIT utterances, 20 % on ten‑times longer inputs, and ultimately 17.6 % PER after the regularization, outperforming the baseline.

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.