Efficient SNR-based subband post-processing for residual noise reduction in speech enhancement algorithms

While current speech enhancement algorithms can significantly reduce background noise, the output speech is commonly unacceptably damaged - a strong penalty for sensitive applications. Alternatively, reducing the aggressiveness leads to more background residual noise - another rejection criterion in practice. In this work, a cost-effective technique for residual noise reduction is presented as a postprocessor for less aggressive enhancement algorithms. The main motivation is to keep their beneficial characteristics, and use the noisy and pre-enhanced signals to remove the remaining noise. The proposed method decomposes pre-enhanced signals into subbands, then performs framewise scaling of the downsampled subband time series based on the estimated Signal-to-Residual-Noise Ratio. Since many popular enhancement algorithms already operate in subbands, the application of the postprocessor is appealing from a computational standpoint. Results show the method consistently reduces background noise, with no further apparent speech damage, as reported by several objective measures and informal listening experiments.