Research On Audio Signal Filtering And Recognition Technology Based On Time-frequency Domain Analysis

In audio and speech recognition systems used for mobile communications, the pollution of variety noises degrades the recognition accuracy significantly due to live recording input from microphones. To reduce the time-varying noise and increase the recognition performance, this thesis utilizes the time-frequency analysis method firstly to separate the signals whose frequency bands overlapped and suppress the noise by the time-frequency(TF) window in the Wigner-Ville Distribution(WVD) domain, with the Fractional Fourier transform(Fr FT). The reconstructed signals of time domain are also obtained with the Wigner Distribution Synthesis techniques(WDST). Simulation results show that the signal separation method is effective to reduce chirp noise. Secondly, with the improved Teager Energy Operator(TEO) and the ratio parameters based on the local variance of the wavelet coefficients, the conventional adapted thresholds are optimized based on wavelet analysis theory and the enhanced speeches are obtained at last with noise reduced. Experimental results demonstrate that under different noise environments, the proposed method increases not only the Signal to Noise Ratio(SNR) but also the Perceptual Evaluation of Speech Quality(PESQ) scores. Finally, research and development are conducted to combine the advantages of the two proposed methods to form a denoising method based on the fractional Fourier transform in Wigner-Ville domain and the energy operator and local variance in wavelet domain. The systemic experiments for signal denoising are also performed at the same time, and results show that the enhancement of speech signal with chirp noise and various noises as in the real world can effectively reduce noise and make the speech recognition more accurately by the fast processing of the combined method.