Statistical Model-Based Tamil Stuttered Speech Segmentation Using Voice Activity Detection

Abstract

Speech is one of the special aid to communicate between humans. A complex speech disorder anointed stuttering is represented by repeated sounds, uttering long syllables or sentences, and blocks of speech. This instinctual speech impairment shows significant issues with the typical fluency and flow of speaking. Voice Activity Detection (VAD) is a vital front-end pre-processing method utilized in numerous speech and signal processing applications to estimate speech presence or absence in short segments of stutter speech intervals. The standard VAD is desirable to extract the stuttered speech signal features through Frame Energy, Zero Crossing Rate and autocorrelation, which detect voiced, Unvoiced or Silence (VUS) signals. Nature-inspired Particle Swarm Optimization (PSO) algorithm is proposed to detect the active Tamil speech using the optimized VAD of the different speakers of Normal Articulate Speech (NAS), Moderate Stutter Speech(MSS) and Severe Stutter Speech (SSS). This primary objective is to update the energy threshold using standard and PSOVAD methods. The proposed PSOVAD performance has been evaluated using objective benchmarks, including Front End Clipping (FEC), Mid-Speech Clipping (MSC), Over Hang (OVER), and Noise Detected as Speech (NDS). According to the experimental findings, PSOVAD can effectively isolate NAS, MSS and SSS into voiced, unvoiced, and silent under low SNR circumstances.