A transformer-based network for speech recognition

In the field of automatic speech recognition (ASR), the noisy audio data and the ambiguity in recognizing homophone lead to the degradation of model performance. In order to address the mentioned problems, a network called DMRS-transformer, a Transformer-based network, is proposed in this study. The proposed DMRS-Transformer mainly consists of two components except for the traditional Transformer network, which are denoising module and Mandarin recognition supplementary module respectively. The denoising module is used for pruning the trivial features caused by the noisy input audio data. The Mandarin recognition supplementary module, short for MRS module, tends to tackle the problem of recognizing Mandarin speech signals which have several homophones. Empirical evaluations have been conducted on two widely used datasets, which are Aishell-1 and HKUST respectively. The experimental results can validate the effectiveness of the proposed DMRS-Transformer network. Compared with the Transformer baseline, the proposed DMRS-Transformer has 0.8% CER improvement and 1.5% CER improvement in these two datasets respectively. © 2023, The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature.