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

Abstract

This paper proposes an efficient memory transformer Emformer for low latency streaming speech recognition. In Emformer, the longrange 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% on test-clean and 5.62% on test-other. Comparing with a strong baseline augmented memory transformer (AM-TRF), Emformer gets 4.6 folds training speedup and 18% relative real-time factor (RTF) reduction in decoding with relative WER reduction 17% on test-clean and 9% on test-other. For a low latency scenario with an average latency of 80 ms, Emformer achieves WER 3.01% on test-clean and 7.09% on test-other. Comparing with the LSTM baseline with the same latency and model size, Emformer gets relative WER reduction 9% and 16% on test-clean and testother, respectively.

Collaborative Training of Acoustic Encoders for Speech Recognition

“…We focus on building models for low latency streaming ondevice speech recognition using the Emformer [2] transducer. Emformer is an efficient extension of the Augmented Memory Transformer (AM-TRF) [1].…”

“…Since we perform experiments for low latency conditions, the center chunk size and look-ahead context size in Emformer are set to 160ms and 40ms, respectively. The algorithmic latency [2] of the acoustic encoders is 120 ms.…”

“…Hence, it has become necessary to build speech recognition models of different sizes to accommodate the wide variety of computational resources. Many automatic speech recognition models [1,2,3] usually are proposed in a few different size configurations, which fall into the relative categories of small, medium, and large. Typically, the models of different sizes are trained independently of each other.…”

“…A se-* Equal Contribution quence transducer model consists of an acoustic encoder, a predictor, and a joiner network. The commonly used acoustic encoders are multi-layered RNN/LSTM [9,10] and Transformers [1,2,3,11]. In this work, we use the low latency streaming Transformer [12] based encoder proposed in the Emformer [2] architecture.…”

“…The commonly used acoustic encoders are multi-layered RNN/LSTM [9,10] and Transformers [1,2,3,11]. In this work, we use the low latency streaming Transformer [12] based encoder proposed in the Emformer [2] architecture.…”