Selflearning Codebook Speech Enhancement

A novel speech enhancement system is presented which exploits a codebook for noise estimation. In contrast to state-of-the-art noise estimators which usually rely on the assumption that the noise signal is only slightly timevarying, codebook approaches allow also non-stationary environments. The basic concept of the proposed codebook noise estimation is a superposition of a scaled speech and noise codebook entry. In order to be independent of a priori noise knowledge, the new estimator is able to learn new noise types online. Training vectors for codebook updates are identified using a speech activity detector (VAD) and a codebook mismatch measure. The VAD is realized as part of the codebook matching. A Wiener filter or any state-of-the-art weighting rule can be applied subsequently for speech enhancement. Experiments confirmed that the new system is able to learn new noise types and provides improved performance compared to state-of-the-art algorithms.