Transformer in action: a comparative study of transformer-based acoustic models for large scale speech recognition applications

TL;DR

This paper compares transformer-based acoustic models, especially Emformer, with traditional LSTM models for large-scale speech recognition, demonstrating significant improvements in word error rates and inference efficiency across various latency scenarios.

Contribution

It provides a comprehensive comparison of transformer-based models with LSTMs in industrial speech recognition tasks, highlighting the advantages of Emformer in accuracy and efficiency.

Findings

Emformer achieves 24-26% relative WERR on low latency voice tasks.

Emformer reduces inference real-time factors by 2-3 times in medium latency scenarios.

Transformer models outperform LSTMs in large-scale speech recognition applications.

Abstract

In this paper, we summarize the application of transformer and its streamable variant, Emformer based acoustic model for large scale speech recognition applications. We compare the transformer based acoustic models with their LSTM counterparts on industrial scale tasks. Specifically, we compare Emformer with latency-controlled BLSTM (LCBLSTM) on medium latency tasks and LSTM on low latency tasks. On a low latency voice assistant task, Emformer gets 24% to 26% relative word error rate reductions (WERRs). For medium latency scenarios, comparing with LCBLSTM with similar model size and latency, Emformer gets significant WERR across four languages in video captioning datasets with 2-3 times inference real-time factors reduction.