Strategies for distant speech recognitionin reverberant environments

Abstract

Reverberation and noise are known to severely affect the automatic speech recognition (ASR) performance of speech recorded by distant microphones. Therefore, we must deal with reverberation if we are to realize high-performance hands-free speech recognition. In this paper, we review a recognition system that we developed at our laboratory to deal with reverberant speech. The system consists of a speech enhancement (SE) front-end that employs long-term linear prediction-based dereverberation followed by noise reduction. We combine our SE front-end with an ASR back-end that uses neural networks for acoustic and language modeling. The proposed system achieved top scores on the ASR task of the REVERB challenge. This paper describes the different technologies used in our system and presents detailed experimental results that justify our implementation choices and may provide hints for designing distant ASR systems.

A generative-discriminative hybrid approach to multi-channel noise reduction for robust automatic speech recognition

“…In particular, dereverberation and beamforming based on linear filtering have been shown to be very effective to improve the robustness of ASR with multi-condition DNN-AMs [12,13]. However, even after performing such filtering, certain amount of residual noise inevitably remains, and substantially limits the improvement of the ASR performance.…”

Integrated Sidelobe Cancellation and Linear Prediction Kalman Filter for Joint Multi-Microphone Speech Dereverberation, Interfering Speech Cancellation, and Noise Reduction

“…Besides iterative LP filter estimation approaches such as [6], [8], [9], [11]- [13], also adaptive approaches based on recursive least squares (RLS) [7], [10], [16], [19] as well as the Kalman filter [14], [15], [17] have been proposed in the past years. In order to reduce noise after dereverberation, multiple-output MCLP has been cascaded with MVDR beamforming in [12], [13], which was seen to be a commonly adopted approach in the 2018 CHiME-5 challenge [23]. In [22], the cascade in [12], [13] is unified.…”

“…We further augment the proposed approach by a spectral Wiener gain post-processor, which is shown to relate to the Kalman filter's posterior state estimate. In order to demonstrate the effectiveness of the ISCLP Kalman filter, we compare against two state-of-the-art approaches -first the previously mentioned alternating Kalman filters in [18], and second a MCLP+GSC Kalman filter cascade, conceptually relating to [12], [13]. As compared to these two reference algorithms, the ISCLP Kalman filter is computationally roughly M 2 times less expensive, where M denotes the number of microphones.…”

Jointly Optimal Dereverberation and Beamforming

“…One issue of the conventional cascade approach [12,13] was that the overall optimality was not guaranteed. The approach performs the optimization separately for dereverberation and beamforming.…”