Novel algorithm for speech segregation by optimized k-means of statistical properties of clustered features

摘要

To simplify the jobs of speaker diarization and speech separation, at first, speech signal should be segregated to two speech formats, dialog and mixture. This paper describes a new algorithm which achieves that first step efficiently. The algorithm is based on Perceptual Linear Predictive feature extraction, optimized k-means and both top-down & bottom-up scenarios. After extracting features of the observation signal, k-means clusters the statistical properties such as variances of the PDF (histogram) of clustered extracted features. k-means is optimized by discounting the worst pattern of clustering step through doing the k-means procedure twice. The feedback loop is necessary for the guiding of the optimized k-means by exploiting the attributes of ordinary k-means. The results of segregation are excellent. The calculated diarization error rate of outputs is very limited.