Variable Quantile Level Based Noise Suppression for Robust Speech Recognition

This paper addresses the issues of single microphone based noise estimation technique for speech recognition in noisy environments. Many researches have been performed on the environmental noise estimation; however, most of them require voice activity detection (VAD) for accurate estimation of noise characteristics. We propose an approach for efficient noise estimation without VAD, aiming at improving the conventional quantile-based noise estimation (QBNE). We fostered the QBNE by adjusting the quantile level according to the relative amount of added noise to the target speech. From the observation that the power spectral density (PSD) of noise is close to the Gaussian distribution, while that of speech is more narrowly populated, the level of additive noise is measured by the selected Gaussianity functions. We compared the proposed method with the conventional QBNE and minimum statistics based method on a simple speech recognition task in various SNR levels. The experimental results show that the proposed method is superior to the conventional methods.