VAD algorithms energy-based and spectral-domain applied in River Plate Castilian

Abstract

Because the English and Castilian have marked acoustic and phonetic differences, this paper shows the study of the effectiveness of different algorithms VAD (Voice Activity Detection) literature, applied to the Castilian, especially riplatense. This article is intended to publicize the results achieved to date. In the first part of the document briefly explained the three implemented methods, namely the autocorrelation function short time (STACF), the average magnitude of the differential junction (FDMA) and the linear prediction coefficients (LPC). Immediately, tests and experiments with BEPPA battery to evaluate the effectiveness of these algorithms VAD will be described. In this step 10 sentences were applied in selected Rioplatense Spanish BEPPA battery of each VAD to detect sound segments, they were used without voice and silence. Immediately, the results obtained in the experimental phase is disclosed, evaluate classifications using the confusion matrix of the 10 phrases in 65 words were about 40 segments of silence. Finally, conclusions and future work are described. Clearly that shows that the algorithms have not been implemented show overall efficiency in detecting voice activity in Spanish of the Rio de la Plata. We also found that the algorithms implemented using linear prediction coefficients show better performance.