---

Higher Order Statistical Analysis of Group Delay Spectrum for Formant Estimation of Noisy Speech Signal

Corresponding Author : Husne Ara Chowdhury (husna-cse@sust.edu)

Keywords : Gaussian noise, Higher-orderstatistics, Group delay, Resonance, Bi-spectrum

Abstract :

The spectral distortion caused by different additive Gaussian noise degrades the speech quality. As a result, it imposes problems when estimating the formant frequency. In this paper, we proposed a technique for 

estimating the formant frequencies from the noisy speech signal. A process is described to suppress the noise of the noise-perverted speech, employing the higher-order statistical (HOS) analysis. After analyzing a few issues related to the magnitude spectrum, we used a minimum phase signal to extract an efficient group delay (GD) signal. Then the higher-order GD spectrum is exploited in place of the higher-order magnitude spectrum. Accurately estimated higher-order GD spectrum emphasizes the vocal-tract system information with the suppressed excitation source information and shows high immunity to noise. The analysis explored throughout this paper exhibits that the higher-order GD spectrum can be an efficient tool for the deconvolution of filter and source in a clean and noisy condition. The proposed method has been tested using the five synthetic vowels showing the

formant frequency estimation error against the pitch variations. The accuracy has also been validated calculating the spectrogram of an utterance from TIMIT dataset and then plotting the formant values on it. The male and

female vowel utterances were verified by the F2–F1 plot. The comparison has been conducted against two state-of-the-art methods. The accuracy of the proposed method is higher than these two methods.

Published on October 23rd, 2024 in Volume 4 Issue 2, Computer Science, Electrical and Electronics