Onset detection algorithm in voice activity detection for Mandarin

Voice activity detection (VAD) is one of the most challenging problems in the field of speech signal processing. The statistical model based VADs have been widely studied in the recent literatures, which usually utilize hangover algorithms to prevent clipping of weak speech tails. However, little attention has been paid on the initial consonants, and non-negligible onset detection errors might be incurred especially when the SNR is low. Since most of the Mandarin syllables start with initial consonants, an onset detection algorithm is proposed in this paper to improve the performance of VAD for Mandarin. Although consonants are mostly noise-like, they produce spectral energy distributed more towards the higher frequencies. To this characteristic, the proposed algorithm makes decision whether the weak-start detection could possibly been dampened by noise based on the posterior SNR of high frequency band, and then it makes correction correspondingly after estimating whether the week-start speech frames mistaken for nonspeech frames exist. It shows that the proposed algorithm achieves a considerable performance improvement. Furthermore, this algorithm is independent of noise type.