Speech parameter extraction in noisy environment using a masking model

To realize a speech recognition system for daily usage, we need to consider the performance degradation due to the surrounding noise. If the surrounding noise is nonstationary, the conventional techniques, such as spectral subtraction or noise injection for templates, are not effective. A new method of speech parameter extraction in noisy environment is proposed as a font-end of a speech recognition system. This method applies a masking model to the observed signal before extracting the LPC parameters. Because the lower level noise components are masked by the speech, the degradation of the extracted parameters is reduced. According to experimental results for isolated word recognition, the improvement using the proposed method is about 4 dB in the SNR for white noise. The total improvement is over 25 dB when this method combined with a noise canceler using an adaptive digital filter