Two-pass Quantile Based Noise Spectrum Estimation

Noise spectrum estimation from a noisy speech signal forms a critical part of such applications as single channel speech enhancement and robust automatic speech recognition (ASR). The two-pass quantile based noise estimation algorithm presented in this paper has the ability to track slow changing non-stationary noise and obtains good estimates for various noise types over a wide range of SNR levels. The essence of the two-pass approach involves first estimating the SNR for each frequency/time point using a fixed quantile value. Based on the estimated SNR, a new quantile level is chosen for each frequency subband. A second noise estimation is then performed. Application of quantile based noise estimation to a single channel speech enhancement system is used to illustrate the ability to suppress noise with little speech distortion.