Speaker Adapted Codebooks for Speech Enhancement

Speech enhancement methods employing a priori information of speech and noise as trained codebooks of speech and noise spectral shapes parametrized by, e.g., linear predictive (LP) coefficients have shown to perform well in non-stationary noise conditions even in single channel mode. Generally, speaker independent (SI) codebooks are employed but for applications such as mobile communication, speaker dependent (SD) code-books are more effective. However, large amount of training data required for generating such SD models is not available in practical application. One way to overcome this limitation is to adapt available SI model to a specific speaker data using smaller amounts of training data incrementally as and when available. In this paper, we investigate the adaption of SI codebook of spectral representation of speech data to a specific target speaker using Vector Quantization Maximum a posteriori (VQ-MAP) algorithm and study its effect on speech enhancement performance. The experimental results indicate that VQ-MAP leads to adapted codebooks which are closer representation of a speaker than SI codebooks and enable better speech enhancement compared to SI models in codebook-based speech enhancement technique.