Speech enhancement using speaker dependent codebooks

Speech enhancement techniques that employ trained codebooks of speech and noise linear predictive coefficients have been previously shown to provide good performance in non-stationary noise environments. The speech codebooks are typically trained using data from a large number of speakers. For noise reduction systems in personal devices such as mobile phones, the use of speaker dependent codebooks is appealing as such devices are typically owned by a single user. It is however not straightforward whether the use of speaker dependent codebooks translates into an observable benefit in terms of noise reduction. Moreover, it is of interest to quantify the improvement obtained over speaker independent codebooks. In this paper, the benefit of using speaker dependent codebooks is analyzed in terms of the resulting improvement in noise reduction when compared to using speaker independent codebooks.