Speech Enhancement Using End-to-End Speech Recognition Objectives

Speech enhancement systems, which denoise and dereverberate distortedsignals, are usually optimized based on signal reconstructionobjectives including the maximum likelihood and minimum meansquare error. However, emergent end-to-end neural methods enableto optimize the speech enhancement system with more applicationorientedobjectives. For example, we can jointly optimize speechenhancement and automatic speech recognition (ASR) only withASR error minimization criteria. The major contribution of this paperis to investigate how a system optimized based on the ASR objectiveimproves the speech enhancement quality on various signallevel metrics in addition to the ASR word error rate (WER) metric.We use a recently developed multichannel end-to-end (ME2E)system, which integrates neural dereverberation, beamforming, andattention-based speech recognition within a single neural network.Additionally, we propose to extend the dereverberation sub networkof ME2E by dynamically varying the filter order in linear predictionby using reinforcement learning, and extend the beamformingsubnetwork by incorporating the estimation of a speech distortionfactor. The experiments reveal how well different signal level metricscorrelate with the WER metric, and verify that learning-basedspeech enhancement can be realized by end-to-end ASR trainingobjectives without using parallel clean and noisy data.