Low Computational Complexity Pitch Based VAD for Dynamic Environment in Hearing Aids

Abstract

This paper presents a low computational complexity and high robust voice activity detection (VAD) algorithm. The algorithm is based on an efficient time-domain pitch detection and harmonic structure discrimination exploiting frequency decomposition by 1/3 octave filter bank defined in ANSI S1.11 standard. In addition, a simple yet efficient phoneme keeper (PK) is adopted for the detection of monosyllable languages, such as Mandarin. Simulation results reveal that the proposed VAD has very robust performance for Mandarin speeches in different environments, even for dynamic SNR (signal to noise ratio) and noise type. Furthermore, in white noise with 0dB SNR, the proposed VAD still has about 90 percent accuracy.