The linear predictive coding (LPC) is a technique that is widely utilized in speech processing, especially in speech spectral envelope modelling, where the aim is to estimate the vocal tract resonances. Consequently, the LPC is used in a vast range of applications, like speech coding, transformation, compression, and speech and speaker recognition. However, with all advances in speech signal processing techniques, the existence of background noise poses a compelling challenge and has a huge impact on the speech quality. This paper proposes an enhancement technique that benefits from the LPC coefficients and exploits them as parameters for an adaptive filter used in speech noise reduction. Unlike the other methods that used the LPC technique for vocal tract spectral envelope estimation or source residual modelling, the proposed method adopts the LPC coefficients to build a frame-dependent, bandpass-like adaptive digital filter that is completely derived from the LPC parameters. This filter will be generated from each speech segment and used to suppress additive noise while preserving real-speech components. The proposed technique is evaluated and compared with two traditional fixed-parameter filters, Wiener and Kalman filters, using speech samples of the TIMIT dataset. The samples are corrupted with additive Gaussian white noise with an input signal-to-noise ratio (SNR) of 20 dB. Practical results show that the performance of the proposed LPC-based adaptive filter in enhancing the speech signal is remarkable with the voiced speech segments, while producing comparable results to Wiener and Kalman filters with unvoiced speech. These points out the effect of frame-dependent LPC-based adaptive filter to deal with nonstationary signals like speech.
Ajel et al. (Fri,) studied this question.