Noise Suppression based on nonnegative matrix factorization for robust speech recognition

This paper presents a novel noise robustness method, nonnegative matrix factorization-based noise suppression (NNS), to enhance the magnitude spectrum of speech signals for better speech recognition performance in noise-corrupted environments. In the presented approach, the clean data and noise in the training set are firstly converted to the spectrograms via short-time Fourier transform (STFT), and the basis spectral matrices of the speech data and noise are learned from the corresponding spectrograms accordingly. Then, the magnitude spectrogram of the noise-corrupted testing data is factorized via the basis matrices of the clean data, and the resulting noise components are alleviated from the original magnitude spectrogram. Finally, the new noisereduced magnitude spectrogram is integrated with the original noisy phase spectrogram and then converted back to a time-domain signal, which is subsequently converted to a sequence of MFCC speech features. By using the presented NNS as a pre-processing stage of the speech recognition system, the obtained recognition accuracy can outperform the MFCC baseline especially at median and low SNR cases. Furthermore, performing NNS on the different sub-band spectrograms can further improve the recognition results relative to the original NNS performing on the full-band spectrogram, indicating that sub-band NNS can produce more robust speech features suitable for noisy speech recognition.