Abstract:
To realize a speech recognition system for our daily life, surrounding noise needs to be considered. Many studies have been done for parameters which are robust in a noisy environment. This paper presents a new parameter which is based on a psychoacoustical rule-base model. First the spectral local peaks of formants are calculated as a basic parameter. By performing this operation frame by frame, the parameter can be represented in a three-dimensional form like a spectrogram. The parameter is then modified by applying psychoacoustical rules to the sequence of peaks, corresponding to a so-called ``auditory stream'' in a speech. The parameter modification utilizes the characteristics of the parameter such as the synchronous onset or offset, the continuity of sequential peaks, and so on. The efficiency of the proposed parameter is evaluated on an isolated word speech recognition system with various SNR. As a result, the improvement due to the proposed parameter is at least 4 dB in SNR for white noise, in comparison with the performance of a LPC-based system.