Accelerator-Aware Training for Transducer-Based Speech Recognition
This addresses the problem of efficient on-device streaming speech recognition for users by reducing latency and maintaining accuracy, though it is incremental as it adapts existing methods to a specific hardware constraint.
The paper tackles the performance degradation of speech recognition models when deployed on neural network accelerators due to full-precision training, by replicating accelerator operations during training to emulate low-precision inference. The result is a 5-7% improvement in engine latency and up to 10% relative reduction in word error rate degradation on a 270K-hour English dataset.
Machine learning model weights and activations are represented in full-precision during training. This leads to performance degradation in runtime when deployed on neural network accelerator (NNA) chips, which leverage highly parallelized fixed-point arithmetic to improve runtime memory and latency. In this work, we replicate the NNA operators during the training phase, accounting for the degradation due to low-precision inference on the NNA in back-propagation. Our proposed method efficiently emulates NNA operations, thus foregoing the need to transfer quantization error-prone data to the Central Processing Unit (CPU), ultimately reducing the user perceived latency (UPL). We apply our approach to Recurrent Neural Network-Transducer (RNN-T), an attractive architecture for on-device streaming speech recognition tasks. We train and evaluate models on 270K hours of English data and show a 5-7% improvement in engine latency while saving up to 10% relative degradation in WER.