Efficient Voice Activity Detection and Speech Enhancement Algorithms based on Spectral Features
Yanna Ma · Institutional Repositories DataBase (IRDB) · 2014
A novel and robust Voice Activity Detection (VAD) algorithm utilizing long-term spectral atness measure (LSFM) and an ecient speech enhancement (SE) algorithm based on modied Wiener ltering method have been proposed in this thesis.The LSFM-based VAD improves speech detection robustness in various noisy environments by employing a low-variance spectrum estimate and an adaptive threshold.The discriminative power of the new LSFM feature is shown by conducting an analysis of the speech/non-speech LSFM distributions.Based on the analysis, we nd that LSFM has the potential to be used as a robust feature for VAD.The proposed LSFM-based VAD algorithm was evaluated under twelve types of noises (eleven from NOISEX-92 and Speech shaped noise) and ve types of signal-to-noise ratio (SNR) in core TIMIT TEST corpus.Comparisons with three modern standardized algorithms (ETSI AMR option 1&2 and ITU-T G.729 AnnexB) demonstrate that our proposed LSFM-based VAD scheme achieved best average accuracy rate.A long-term signal variability (LTSV)-based VAD scheme is also compared with our proposed method.The results show that our proposed algorithm outperforms it for most of the noises considered including dicult noises like Machine gun noise and Speech babble noise.After the introduction of the LSFM-based VAD, we continue to show the proposed ecient SE algorithm.It utilized constraints to the Wiener gain function in which the wavelet thresholded multitaper spectrum was taken as the clean spectrum for the constraints.The proposed algorithm was evaluated under eight types of noises and ii seven SNR levels in NOIZEUS database and was predicted by the composite measures and the SNR LOSS measure to improve subjective quality and speech intelligibility in various noisy environments.Comparisons with two other algorithms (KLT and WT) demonstrate that in terms of signal distortion, overall quality and the SNR LOSS measure, our proposed constrained SE algorithm outperforms the KLT and WT schemes for most conditions considered.