A Method of Real-Time Non-uniform Speech Stretching

Developed method of real-time non-uniform speech stretching is presented. The proposed solution is based on the well-known SOLA algorithm (Synchronous Overlap and Add). Non-uniform time-scale modification is achieved by the adjustment of time scaling factor values in accordance with the signal content. Dependently on the speech unit (vowels/consonants), instantaneous rate of speech (ROS), and speech signal presence, values of the scaling factor are selected. This provides as low as possible difference in the duration of the input and output signal and high naturalness and quality of the modified speech. In the experimental part of the paper accuracy of the proposed ROS estimator is examined. Quality of the speech stretched using the proposed method is assessed in the subjective tests.