Line spectral frequency-based noise suppression for speech-centric interface of smart devices

This paper proposes a noise suppression technique for speech-centric interface of various smart devices. The proposed method estimates noise spectral magnitudes from line spectral frequencies (LSFs), using the observation that adjacent LSFs correspond to peak frequencies of spectrum, whereas isolated LSFs are close to flattened valley frequencies retaining noise components. Over a course of segmented time frames, the logarithms of spectral magnitudes at respective LSFs are computed, and their distribution is then modeled by the Rayleigh probability density function. The standard deviation from the Rayleigh function approximates the noise spectral magnitude. The model is updated at every frame in an online manner so that it can deal with real-time inputs. Once the noise spectral magnitude is estimated, a time-domain Wiener filter is derived for the suppression of the estimated noise spectral magnitude, and this is then applied to the input noisy speech signals. Our proposed approach operates well on most smart devices owing to its low computational complexity and real-time implementation. Speech recognition experiments, conducted to evaluate the proposed technique, show that our method exhibits superior performance, with less distortion of original speech, when compared to conventional noise suppression techniques.