Modeling the effects of single-microphone noise-suppression

Many different algorithms have been proposed for single-microphone noise suppression. Comparing algorithms can be difficult, however, since different studies use different speech stimuli, different types of noise, and different signal-to-noise ratios (SNRs). This paper proposes the HASPI intelligibility and HASQI quality metrics as effective tools to measure and compare the performance of noise-suppression algorithms. The stimuli are sentences presented in multi-taker babble at SNRs ranging from −20 to +40 dB. Six different noise-suppression algorithms are compared: Wiener, power, and magnitude spectral subtraction, a smoothed gain-vs-SNR approach that limits the maximum attenuation, binary mask processing, and auditory masked transform noise suppression. Exact restoration of the speech envelope and no processing provide reference conditions. Also investigated are the impact of differing amounts of knowledge about the noise signal and the effects of impaired versus normal hearing. In general, the algorithms that worked best for normal hearing were also most effective for a simulated moderate hearing loss. The results indicate that many algorithms provide high predicted intelligibility under ideal conditions where the speech and noise power are known for every segment within each frequency band, but that the performance degrades to that measured for no processing when the local noise power is replaced by the power averaged over all of the segments within the band. The upper bound to algorithm performance is shown to be exact restoration of the noisy speech envelope to match that of the noise-free signal.