In this paper, a Mel-scaled AR (Mel-AR) model based VAD is presented, where likelihood ratio measure is used to classify the input speech frames as speech/non-speech segments. The Mel-AR model parameters have been estimated on the linear frequency scale from the input speech signal without applying bilinear transformation. This has been done by employing a first-order all-pass filter rather than unit delay. The performance of the proposed VAD is evaluated on Aurora-2 database by measuring FAR and FRR. The equal false rate (EFR) at the crossover point is also presented as a merit of VAD. In addition, the performance of the proposed VAD in speech recognition is verified by incorporating it with a Mel-Wiener filter for MLPC based noisy speech recognition.

This paper deals with a wavelet domain Wiener filter to estimate enhanced Mel-LPC spectra in presence of additive noises. In this implementation, Daubechies 4 (db4) wavelet function has been used as mother wavelet which enables a fast computation and decomposition using perfect reconstruction of filterbank. To implement the filter, noise is estimated from the initial 20 frames of input speech signal without applying any voice activity detection (VAD) system. In the proposed system, filtering is done in wavelet domain using Wiener gain. After filtering, inverse wavelet transform is applied to obtain enhanced time domain speech signal. Using this enhanced speech signal Mel-LP cepstral coefficients are calculated as speech feature. The proposed system is evaluated on Aurora-2 database and it has been found that the Wiener filter improves the overall word accuracy from 58.66 to 75.88% and the average Aurora-2 relative improvement has been found to be 42.50% for test set A.