Robust Speech Recognition System for Communication Robots in Real Environments

摘要

The application range of communication robots could be widely expanded by the use of an automatic speech recognition (ASR) system with improved robustness for noise and for speakers of different ages. In this paper, we describe an ASR system which can robustly recognize speech by adults and children in noisy environments. We evaluate the ASR system in a communication robot placed in a real noisy environment. Speech is captured using a twelve-element microphone array arranged in the robot chest. To suppress interference and noise and to attenuate reverberation, we implemented a multi-channel system consisting of an outlier-robust generalized sidelobe canceller (RGSC) technique and a feature-space noise suppression using MMSE criteria. Speech activity periods are detected using GMM-based end-point detection (GMM-EPD). Our ASR system has two decoders for adults' and children's speech. The final hypothesis is selected based on posterior probability. We then assign a generalized word posterior probability (GWPP)-based confidence measure to this hypothesis, and if it is higher than a threshold, we transfer it to a subsequent dialog processing module. The performance of each step was evaluated for adults' and children's speech, by adding different levels of real environment noise recorded in a cafeteria. Experimental results indicated that our ASR system could achieve over 80 % word accuracy in 70 dBA noise. Further evaluation of adult speech recorded in a real noisy environment resulted in 73 % word accuracy