Noisy speech recognition with microphone array steering and Fourier/wavelet spectral subtraction
Yuki Denda, Takanobu Nishiura, Hideki Kawahara · IEEE Signal Processing Workshop on Statistical Signal Processing · 2003
It is very important to capture distant-talking speech with high quality for teleconferencing systems or voice-controlled systems. For this purpose, microphone array steering and Fourier spectral subtraction, for example, are ideal candidates. A combination technique using both microphone array steering and Fourier spectral subtraction has also been proposed to improve performance. However, it is difficult for the conventional approach to reduce non-stationary noise, although it is easy to robustly reduce stationary noise. To cope with this problem, we propose a new combination technique with microphone array steering and Fourier/wavelet spectral subtraction. Wavelet spectral subtraction promises to effectively reduce non-stationary noise, because the wavelet transform admits a variable time-frequency resolution on each frequency band. As a result of evaluation experiments in a real room, we confirmed that the proposed combination technique provides better performance of the ASR (automatic speech recognition) and NRR (noise reduction rate) than the conventional combination technique in a directional and a diffused noise environment.