English

Emformer: Efficient Memory Transformer Based Acoustic Model For Low Latency Streaming Speech Recognition

Sound 2021-01-01 v4 Computation and Language Machine Learning Audio and Speech Processing

Abstract

This paper proposes an efficient memory transformer Emformer for low latency streaming speech recognition. In Emformer, the long-range history context is distilled into an augmented memory bank to reduce self-attention's computation complexity. A cache mechanism saves the computation for the key and value in self-attention for the left context. Emformer applies a parallelized block processing in training to support low latency models. We carry out experiments on benchmark LibriSpeech data. Under average latency of 960 ms, Emformer gets WER 2.50%2.50\% on test-clean and 5.62%5.62\% on test-other. Comparing with a strong baseline augmented memory transformer (AM-TRF), Emformer gets 4.64.6 folds training speedup and 18%18\% relative real-time factor (RTF) reduction in decoding with relative WER reduction 17%17\% on test-clean and 9%9\% on test-other. For a low latency scenario with an average latency of 80 ms, Emformer achieves WER 3.01%3.01\% on test-clean and 7.09%7.09\% on test-other. Comparing with the LSTM baseline with the same latency and model size, Emformer gets relative WER reduction 9%9\% and 16%16\% on test-clean and test-other, respectively.

Keywords

Cite

@article{arxiv.2010.10759,
  title  = {Emformer: Efficient Memory Transformer Based Acoustic Model For Low Latency Streaming Speech Recognition},
  author = {Yangyang Shi and Yongqiang Wang and Chunyang Wu and Ching-Feng Yeh and Julian Chan and Frank Zhang and Duc Le and Mike Seltzer},
  journal= {arXiv preprint arXiv:2010.10759},
  year   = {2021}
}

Comments

5 pages, 2 figures, submitted to ICASSP 2021

R2 v1 2026-06-23T19:30:36.597Z