An efficient joint training model for monaural noisy-reverberant speech recognition

Noise and reverberation can seriously reduce speech quality and intelligibility, affecting the performance of downstream speech recognition tasks. This paper constructs a joint training speech recognition network for speech recognition in monaural noisy-reverberant environments. In the speech enhancement model, a complex-valued channel and temporal-frequency attention (CCTFA) are integrated to focus on the key features of the complex spectrum. Then the CCTFA network (CCTFANet) is constructed to reduce the influence of noise and reverberation. In the speech recognition model, an element-wise linear attention (EWLA) module is proposed to linearize the attention complexity and reduce the number of parameters and computations required for the attention module. Then the EWLA Conformer (EWLAC) is constructed as an efficient end-to-end speech recognition model. On the open source dataset, joint training of CCTFANet with EWLAC reduces the CER by 3.27%. Compared to other speech recognition models, EWLAC maintains CER while achieving much lower parameter count, computational overhead, and higher inference speed.