Speech recognition in a noisy environment using a noise reduction neural network and a codebook mapping technique

Three methods are presented for speech recognition in a noisy environment using a noise reduction neural network, a codebook mapping technique, and a combination of these methods. Noisy speech data was generated artificially by adding computer room noise or pink noise to speech. First the codebook mapping technique was tested using artificial pink noise in HMM-LR Japanese phrase recognition experiments. As a result, it was confirmed that the codebook mapping technique has the ability to reduce some noise. Next, the methods were tested on a phoneme recognition task for /b, d, g/ using HMMs in actual computer room noise. It was found that the recognition rates with the noise reduction neural network, the codebook mapping, and the combination method are 59.9%, 58.3%, and 62.3%, respectively. These recognition rates were higher than the recognition rate (43.8%) without noise reduction. The result shows that the three methods are effective recognition methods for noise-corrupted speech