Adaptive filtering using harmonic structure of voiced speech for reducing nonstationary known noise

Abstract

Proposed is an effective method for reducing nonstationary known noise. The objective of this research is to develop a scheme of preprocessing for speech recognition that keeps the same speech recognition rate even in a worse acoustic environment and to realize a TV control system using speech recognition. The basic idea of the proposed method is to estimate the frequency spectrum of noise including sounds from the TV itself and to remove noise components from the frequency spectrum of the received signal. A transfer function from the TV set to the microphone is calculated in an iterative manner estimating the noise signal at the microphone. Traditional ANC techniques do not directly use spectral features such as harmonic structure or fundamental frequency, for example. As the proposed method uses harmonic structure of vocalic segments in command speech, it is expected to have an advantage compared with the traditional adaptive methods. Results of evaluation using speech recognition show that the proposed method significantly improves speech recognition rate compared with conventional spectral subtraction. [Work supported by Knowledge Cluster Project, MEXT, and by Academic Frontier Project Doshisha University.]