Large Margin Training for Attention Based End-to-End Speech Recognition

By Peidong Wang, Jia Cui, Chao Weng, Dong Yu. 2019

8 citations recorded by the source corpus.

A recent research trend in speech recognition is to jointly optimize acoustic and language models within a unified end-to-end frame-work. Due to the difficulty in training such models, a variety of optimization methods have been investigated. In this study, we propose a novel large marge training scheme for attention based end-to-end speech recognition. Using only one hypothesis during training, it can achieve comparable performance to the minimum Bayes risk based minimum word error rate (MWER) criterion. When character-level edit distance is applied to both methods, large margin is even consistently better than MWER. The proposed training scheme is also a more succinct formulation of the large margin concept. It keeps the original model structure and is thus easier to apply.

Subfields: Artificial Intelligence, Speech & Audio, Signal Processing