Corpus based reconstruction of speech degraded by wind noise

This contribution addresses the problem of enhancing a speech signal which is degraded by wind noise. The characteristic that wind noise signals are sparse in time and frequency is exploited in a way that only time-frequency regions that are determined as degraded are enhanced. In these regions of the noisy signal, a process is applied to reconstruct the clean speech data. This is realized by a separation of the noisy speech signal into an autoregressive filter representing the human vocal tract and its excitation signal. The clean filter coefficients of the former are estimated using a pre-trained codebook. A pitch cycle taken from clean speech is adapted to reconstruct the excitation of noisy speech segments.