Noise Characterization and Classification for Background Estimation

Characterization and classification of noise in the noisy speech is an important part of noisy speech processing in all related fields of application for speech technology. In this paper, problem of noise characterization and categorization of various environmental noises in the noisy speech of different SNR´s (Signal to Noise Ratio) has been handled. This paper purposes a technique for estimation of noise level and classifying the noises in a noisy speech into different categories of noises for background estimation. The technique detects non-speech regions in the noisy speech for SNR estimation and noise classification purpose. Three dimensional representation of noise is obtained using technique, which discriminates visual imprints of different noises in most optimum way in the 2D plane. For classification of test noise, prominent features from the 2D representation of noise are selected and classified using minimum hamming distance classifier using reference standard noise. Two dimensional second order derivatives of FFT magnitude matrix w.r.t. time and frequency are computed from the extracted pause regions. The technique is computationally simple and efficient and gives very high level of accuracy in the noisy speech with segmental SNR range of 20 dB to -20 dB. The test noises are categorized into white, babble, factory, vehicular and channel noises. The efficacy of technique developed is dependent on the limitations of the particular speech and pause detection algorithm used, particularly at moderate value of SNR. At lower value of SNR no such limitations exists and accuracy goes up to even 100%.