A Computationally Efficient Mel-Filter Bank VAD Algorithm for Distributed Speech Recognition Systems

Abstract

This paper presents a novel computationally efficient voice activity detection (VAD) algorithm and emphasizes the importance of such algorithms in distributed speech recognition (DSR) systems. When using VAD algorithms in telecommunication systems, the required capacity of the speech transmission channel can be reduced if only the speech parts of the signal are transmitted. A similar objective can be adopted in DSR systems, where the nonspeech parameters are not sent over the transmission channel. A novel approach is proposed for VAD decisions based on mel-filter bank (MFB) outputs with the so-called Hangover criterion. Comparative tests are presented between the presented MFB VAD algorithm and three VAD algorithms used in the G.729, G.723.1, and DSR (advanced front-end) Standards. These tests were made on the Aurora 2 database, with different signal-to-noise (SNRs) ratios. In the speech recognition tests, the proposed MFB VAD outperformed all the three VAD algorithms used in the standards by 14.19% relative (G.723.1 VAD), by 12.84% relative (G.729 VAD), and by 4.17% relative (DSR VAD) in all SNRs

rVAD: An unsupervised segment-based robust voice activity detection method

“…Results for the G.729, G.723.1 and MFB VAD methods are cited from [17], results for the LTSV and GMM-NLSM methods are from [27], and results for the DSR-AFE and variable frame rate (VFR) methods are from [3]. The comparison in this table is conducted in terms of frame error rate (FER) since results of LTSV and GMM-NSLM are only available in terms of FER.…”

“…The comparison in this table is conducted in terms of frame error rate (FER) since results of LTSV and GMM-NSLM are only available in terms of FER. Note that the identical experimental settings and labels are used across [3,17,27] and the present work, so the comparison is valid.…”

“…VAD is a binary classification problem involving both feature extraction and classification. Various speech features can be found in literature such as energy, zero-crossing rate, harmonicity [13], perceptual spectral flux [14], Mel-frequency cepstral coefficient (MFCC) [11], power-normalized cepstral coefficients (PNCCs) [15], entropy [16], Mel-filter bank (MFB) outputs [17] and a posteriori signal-to-noise ratio (SNR) weighted energy distance [3,18]. For the purpose of modelling and classification, popular techniques include Gaussian models [19], Gaussian mixture models (GMM) [2,14,20], super-Gaussian models and their convex combination [9], i-vector [15,21], decision trees [22], support vector machines [23], and neural network models (including deep models) [5,7,24].…”

Long-term auto-correlation statistics based voice activity detection for strong noisy speech

“…V AD is an important frontend in many speech-related applications, such as mobile communication system [1], echo cancellation [2], speech enhancement [3], speech coding [4], automatic speech recognition [5], etc. The accuracy of V AD is quite critical to the overall performance of those applications.…”

Influence of Hangover and Hangbefore Criteria on Automatic Speech Recognition

“…Second, automatic speech recognition systems no longer have to work on those parts of the signal, which contain only noise, the benefit of which is a faster speech recognition process. The second advantage is based on a frame dropping strategy [1,2,3]. In addition to a speed-up of the speech recognition process the frame dropping strategy can also increase its accuracy, if the VAD algorithm gives the right information, about which frame contains only noise and is, therefore, excluded from the further speech recognition process.…”