Latent Sequence Decompositions

TL;DR

This paper introduces the Latent Sequence Decompositions framework, which decomposes sequences into variable length units conditioned on input and output, improving speech recognition accuracy.

Contribution

The paper proposes a novel LSD framework with training and decoding algorithms, demonstrating significant improvements on speech recognition tasks.

Findings

Achieved 12.9% WER on WSJ with LSD model

Reduced WER to 9.6% using convolutional encoder

Outperformed character baseline in speech recognition

Abstract

We present the Latent Sequence Decompositions (LSD) framework. LSD decomposes sequences with variable lengthed output units as a function of both the input sequence and the output sequence. We present a training algorithm which samples valid extensions and an approximate decoding algorithm. We experiment with the Wall Street Journal speech recognition task. Our LSD model achieves 12.9% WER compared to a character baseline of 14.8% WER. When combined with a convolutional network on the encoder, we achieve 9.6% WER.