Attention-based Contextual Language Model Adaptation for Speech Recognition

TL;DR

This paper presents an attention-based method for adapting language models in speech recognition by incorporating contextual information, leading to significant perplexity reductions on a large voice assistant dataset.

Contribution

The paper introduces an attention mechanism for neural language model adaptation that effectively utilizes utterance-level contextual data in speech recognition.

Findings

Reduces perplexity by 7.0% on a large dataset.

Improves perplexity by 9.0% on long-tail utterances.

Outperforms existing contextual language models by over 2.8%.

Abstract

Language modeling (LM) for automatic speech recognition (ASR) does not usually incorporate utterance level contextual information. For some domains like voice assistants, however, additional context, such as the time at which an utterance was spoken, provides a rich input signal. We introduce an attention mechanism for training neural speech recognition language models on both text and non-linguistic contextual data. When applied to a large de-identified dataset of utterances collected by a popular voice assistant platform, our method reduces perplexity by 7.0% relative over a standard LM that does not incorporate contextual information. When evaluated on utterances extracted from the long tail of the dataset, our method improves perplexity by 9.0% relative over a standard LM and by over 2.8% relative when compared to a state-of-the-art model for contextual LM.