Does speech enhancement work with end-to-end ASR objectives? Experimental analysis of multichannel end-to-end ASR

Recently we proposed a novel multichannel end-to-end speech recognition architecture that integrates the components of multichannel speech enhancement and speech recognition into a single neural-network-based architecture and demonstrated its fundamental utility for automatic speech recognition (ASR). However, the behavior of the proposed integrated system remains insufficiently clarified. An open question is whether the speech enhancement component really gains speech enhancement (noise suppression) ability, because it is optimized based on end-to-end ASR objectives instead of speech enhancement objectives. In this paper, we solve this question by conducting systematic evaluation experiments using the CHiME-4 corpus. We first show that the integrated end-to-end architecture successfully obtains adequate speech enhancement ability that is superior to that of a conventional alternative (a delay-and-sum beamformer) by observing two signal-level measures: the signal-todistortion ratio and the perceptual evaluation of speech quality. Our findings suggest that to further increase the performances of an integrated system, we must boost the power of the latter-stage speech recognition component. However, an insufficient amount of multichannel noisy speech data is available. Based on these situations, we next investigate the effect of using a large amount of single-channel clean speech data, e.g., the WSJ corpus, for additional training of the speech recognition component. We also show that our approach with clean speech significantly improves the total performance of multichannel end-to-end architecture in the multichannel noisy ASR tasks.