Clustering based voiced-unvoiced-silence detection in speech using temporal and spectral parameters

This paper reports automatic segmentation of voiced, unvoiced and silence portion of speech on TIM IT data base. Waveform and frequency domain parameters are used to form multi dimensional feature space. Short time energy threshold of unvoiced segment is used to separate out silence or background from speech. The Gaussian similarity function based spectral clustering is used to find error performance of voiced/unvoiced (V/UV) portion of the speech. The classification accuracy of V/UV is measured and the result is compared with the other techniques available in the literatures. The proposed technique provides at least 98.3% V/UV detection accuracy.