arXiv
We design an online end-to-end speech recognition system based on Time-Depth Separable (TDS) convolutions and Connectionist Temporal Classification (CTC). The system has almost three times the throughput of a well tuned hybrid ASR baseline while also having lower latency and a better word error rate. We improve the core TDS architecture in order to limit the future context and hence reduce latency while maintaining accuracy. Also important to the efficiency of the recognizer is our highly optimized beam search decoder. To show the impact of our design choices, we analyze throughput, latency and accuracy and also discuss how these metrics can be tuned based on the user requirements.
Donglai Xiang, Timur Bagautdinov, Tuur Stuyck, Fabian Prada, Javier Romero, Weipeng Xu, Shunsuke Saito, Jingfan Guo, Breannan Smith, Takaaki Shiratori, Yaser Sheikh, Jessica Hodgins, Chenglei Wu
Ludwig Sidenmark, Mark Parent, Chi-Hao Wu, Joannes Chan, Michael Glueck, Daniel Wigdor, Tovi Grossman, Marcello Giordano
Simon Vandenhende, Dhruv Mahajan, Filip Radenovic, Deepti Ghadiyaram
