This paper combined the HMM composition method with a highly efficient noise reduction method to create a robust speech recognition technique for noisy networks. The HMM composition method, i.e. noise and voice composition(NOVO), is well-known as an effective noise adaptation method that can improve speech recognition performance in noisy networks. However, recognition performance of noise adaptation methods like NOVO, is not sufficient if S/N is low, because the speech features are buried in the noise.
Methods to raise the S/N of the observed signals by using noise reduction such as the SS(Spectral Subtraction) and MWF(modified Wiener Filter) method can be used. These methods, however, are not able to remove noise completely, and they create new problems in that insufficient or excessive reduction processing leads to remaining noise or speech distortion, respectively. So, we modified the conventional Spectral Subtraction method, and it was named by MSS(modified Spectral Subtraction).
To deal with the low S/N, we propose noise reduction by MSS-NOVO, a combination of the NOVO method and MSS. This method realizes strong noise reduction with low speech distortion, and the remaining noise is handled by the NOVO method.
In order to evaluate the proposed method, we compared with other method, i.e. SS, MWF, NOVO, SS-NOVO, MWF-NOVO. The speech data used are KLE 452 database. To set up the noisy speech for test, we add the speech signals to the white noise with different SNRs range from 0dB to 20dB, at 5dB intervals.
Form the results, we can see that in noisy networks, SS, MWF, NOVO, SS-NOVO, MWF-NOVO method can improve the speech recognition accuracy. But the improvement is quite limited, especially under 10dB and improvement decreases as the decrease of the SNR. The other way round, the proposed MSS method have more improvement than other conventional methods in 0dB~10dB environment. Especially, MSS-NOVO method have strong improvement in low S/N environment. we can get the improvement results compared with the NOVO and MWF-NOVO method that were 30%, 17%, 3% in 0dB, 5dB, 10dB environment, respectively.
Thus, the evaluation experiments showed that the proposed method is effective in noisy networks.