LSP-based acoustic echo cancellation and noise reduction
MetadataShow full item record
This thesis employs line spectral pairs (LSP) for acoustic echo cancellation (AEC) and noise reduction. LSP coefficients possess several important properties which make them attractive for use in speech enhancement and speech coding techniques, such as adaptive Wiener filtering and QCELP vocoder. Since LSP parameters have been computed for transmission in QCELP, these coefficients can be applied to an acoustic echo canceler. In conventional acoustic echo cancellation, the system attempts to model the acoustic echo path with far-end talking exclusively. During double-talk, the adaptive filter coefficients are frozen to prevent the echo canceler from being affected by the near-end?s signal. In order to obtain the best near-end speech quality, two conditions have to be assumed that (1) the echo path will not change during double-talk; (2) the adaptive filter has already converged to optimal value. These two conditions are not always satisfied, especially in a moving automobile. If the phase of the echo path changing or a impulse response shift occurs, the near-end speech will be seriously degraded. Therefore, a new echo canceler is proposed in which an adaptive Wiener filter is added during double-talk. Inter- and intra-frame spectral constraints are applied to LSP coefficients to ensure convergence to reasonable values and hence lead to improved speech quality. The basis of these evaluation involved comparisons of results from conventional schemes to from the new algorithm with near-end speech degraded by acoustic echo. Noticeable quality improvement can be observed by the new algorithm with echo path changing. Furthermore, a new echo canceler is designed to integrate acoustic echo cancellation with noise reduction. The algorithm is evaluated over different signal-to-interference-and-noise ratio condition. Different measure methods (Itakura distance, segment SNR and informal listening tests) are used to verify it. All these tests show that those obvious speech quality improvement can be achieved by this algorithm.