Corrective algorithm for esophageal voice cycle detection

In this paper a corrective solution for more accurate esophageal voice cycle marks detection is presented. While the algorithm developed in a previous work adapts the MDVP voice period marks to esophageal voice, it shows lacks when analyzing unsteady |a| vowel samples and with high levels of noise. The proposed solution is aimed to iteratively correct the parameters (nmin and threshold) used in the previous algorithm. The corrective system filters the noise of voice recordings and then uses the output of the previous algorithm to calculate the most reliable fundamental frequency, based on the most stable part of the given signal. The results were acquired by comparing the pitch values of the previous algorithm with and without the proposed solution and the calculating the error against the real value. The corrective solution presents a considerable improvement in the accuracy of the voice cycle detection algorithm by reducing the standard deviation of the error dataset from 4.591 to 0.593 and the mean error value from 2.459 down to 0.370. Furthermore, it reduces up to a 54% all errors above 1%.