SNMF Based Speech Denoising with Wavelet Decomposed Signal Selection

This paper proposed the model of non-negative matrix factorization (NMF) with the effect of digital wavelet decomposition in speech denoising. Sparse NMF has been used over magnitude spectrogram of speech signal to find the basis vectors of training and weights of test signal. The results are validating the effect of wavelet decomposition on the performance. To test the algorithm, TIMIT data for speech signal database and Noisex92 data for noise database was used. The performance measurement has been taken in terms of signal to distortion ratio (SDR), signal to artifacts ratio (SAR), signal to interference ratio (SIR), perceptual evaluation of speech quality (PESQ) and short-time objective intelligibility measure (STOi). Here, the separation of speech signal from the noisy signal has been performed with and without prior knowledge of noise. Results are compared with the existing algorithms. Proposed model has shown improvements over the existing models in both conditions.