Adaptive framing based similarity measurement between time warped speech signals using Kalman filter

Abstract

Similarity measurement between speech signals aims at calculating the degree of similarity using acoustic features that has been receiving much interest due to the processing of large volume of multimedia information. However, dynamic properties of speech signals such as varying silence segments and time warping factor make it more challenging to measure the similarity between speech signals. This manuscript entails further extension of our research towards the adaptive framing based similarity measurement between speech signals using a Kalman filter. Silence removal is enhanced by integrating multiple features for voiced and unvoiced speech segments detection. The adaptive frame size measurement is improved by using the acceleration/deceleration phenomenon of object linear motion. A dominate feature set is used to represent the speech signals along with the pre-calculated model parameters that are set by the offline tuning of a Kalman filter. Performance is evaluated using additional datasets to evaluate the impact of the proposed model and silence removal approach on the time warped speech similarity measurement. Detailed statistical results are achieved indicating the overall accuracy improvement from 91 to 98% that proves the superiority of the extended approach on our previous research work towards the time warped continuous speech similarity measurement.