Continuous speech recognition (CSR) experiments were performed to evaluate the performance of the CSCC method to recognize Japanese sentense speech in noisy environment using an microphone array.
We used ``IPA-testset'' consisting of 100 sentences each uttered by male and female speakers excerpted from ASJ-JNAS corpus of read newspaper articles as the test set. Other 10 sentence utterances from the same database were used as 1 to 5 interfering speech noises with a signal-to-noise ratio of 10dB per noise.
Input speech data were analyzed with a 25-mS frame length and 10-mS frame shift. 12-order MFCCs, their MFCCs and log-power were used as acoustic feature vector. Using ``Julius3.3p3''[5] as the speech recognition platform, word accuracy was evaluated as the measure of speech recognition performance.