Noise Reduction Based on Soft Masks by Incorporating SNR Uncertainty in Frequency Domain

作者：Saleem Nasir^*; Irfan Muhammad

来源：Circuits, Systems, and Signal Processing, 2018, 37(6): 2591-2612.

DOI：10.1007/s00034-017-0684-5

• Ideal binary mask
• Noise reduction
• Residual noise
• Soft masking
• Intelligibility
• Quality
• Threshold

The binary mask approach has been studied recently to reduce the background noise and improve the speech intelligibility and quality in the noisy surroundings. This mask is usually applied at the time-frequency illustration of a noisy speech and discards portions of a speech below a signal-to-noise-ratio (SNR) threshold, whereas allowing others to pass over intact. The threshold, however, is normally very low, and considerable residual noise would exist. Moreover, the precise estimate of local instantaneous SNR in practical applications is a difficult task. By modeling the local instantaneous SNR as Fisher-Snedecor distributed random variable, the soft masks for noise reduction are derived by incorporating SNR uncertainty in the frequency domain. Instead of finding a different method to estimate the local instantaneous SNR, the probability of local instantaneous SNR is computed higher than the threshold. The results indicated that soft masks yielded significantly better speech quality in terms of speech distortion and residual noise.

• 出版日期2018-6