In this paper, by using the properties of the higher order statistics (HOS) of speech and noise signals, we develop an improved voice activity detection (VAD) scheme. The proposed scheme employs the logarithm of the kurtosis of the LPC residual of a speech signal and is shown to be more effective and efficient in detecting active speech in medium to low signal-to-noise ratio (SNR) conditions without being unduly affected by the variations in the signal energy. To overcome the inability of the HOS in detecting unvoiced speech, another metric (the low band to full band energy ratio) is introduced. Depending on the estimated mean SNR, the proposed scheme works adaptively in two modes: a simple mode using only the SNR, and an enhanced mode using the HOS, the low band to full band energy ratio and the SNR. This scheme is capable of avoiding unnecessary computations, while maintaining the same performance as that working only in the enhanced mode. Simulations results are presented to demonstrate the effectiveness of the proposed voice activity detection scheme.

An improved voice activity detection using higher order statistics

In many speech signal processing applications, voice activity detection (VAD) plays an essential role for separating an audio stream into time intervals that contain speech activity and time intervals where speech is absent. Many features that reflect the presence of speech were introduced in literature. However, to our knowledge, no extensive comparison has been provided yet. In this article, we therefore present a structured overview of several established VAD features that target at different properties of speech. We categorize the features with respect to properties that are exploited, such as power, harmonicity, or modulation, and evaluate the performance of some dedicated features. The importance of temporal context is discussed in relation to latency restrictions imposed by different applications. Our analyses allow for selecting promising VAD features and finding a reasonable trade-off between performance and complexity.

/pdf/features-for-voice-activity-detection-a-comparative-analysis-1gveo0tf7a.pdf

Features for voice activity detection: a comparative analysis

Computationally efficient speech extraction algorithms have significant potential economic benefit, by automating an extremely tedious manual process. Previously, algorithms which discriminate between speech and one specific other signal type have been developed, and often fail when the specific non-speech signal is replaced by a different signal type. Moreover, several such signal specific discriminators have been combined in order to tackle the general speech vs. non-speech discrimination problem, with predictable negative results. When the number of discriminating features is large, compression methods such as principal components have been applied to reduce dimension, even though information may be lost in the process. In this paper, graphical tools are applied to determine a set of features which produce excellent speech vs. nonspeech clustering. This cluster structure provides the basis for a general speech vs. non-speech discriminator, which significantly outperforms the TALKATIVE speech extraction algorithm.

A multivariate speech activity detector based on the syllable rate

There are many applications for which it is desireable to reliably detect the presence of speech. Examples of these applications are speech compression, voice activated devices and machine speech recognition. In this paper, a method of speech detection is developed which uses a frequency-domain pitch-based signal-to-noise ratio (SNR) estimate. This method takes full advantage of the spectral structure of pitch, which is the primary speech excitation function. The primary output of the detection algorithm is a decision that speech is present or not present. In addition, the algorithm provides an estimate of the speech SNR which may be used to estimate signal quality. This SNR estimate is important for applications such as estimating the reliability of machine-based recognition processes. Additional advantages of this method are that it is independent of signal gain and it works well under adverse conditions such as poor SNR and in the presence of interference. A by-product of the pitch-based detection process is a method for automatic recovery of frequency offset of mistuned analog speech. Mistuning is a condition which can arise in the demodulation of single-side-band amplitude-modulated (SSB-AM) speech if the precise carrier is not used in the demodulation process. This can cause severe problems since speech becomes nearly unintelligible if it is mistuned more than 100 Hz. The methods presented here use a double complex correlation of the complex speech spectrum to recover the carrier offset. This process provides significantly better resolution than more conventional correlation processes based on the speech power- spectrum.© (1995) COPYRIGHT SPIE--The International Society for Optical Engineering. Downloading of the abstract is permitted for personal use only.

Pitch based methods for speech detection and automatic frequency recovery

There are four sorts of men: He who knows not and knows not that he knows not: He is a fool—shun him; He who knows not and knows that he knows not: He is simple—teach him; He who knows and knows not that he knows: He is asleep—wake him; He who knows and knows that he knows: He is wise—follow him. —Arabic proverb Our goal is to develop a robust anomaly detector that can be incorporated into pattern recognition systems that may need to be taught, but will never be shunned. The ability to know what we do not know is a concept often overlooked when developing classifiers to discriminate between different types of normal data in controlled experiments. We believe that an anomaly detector should be used to produce warnings in real applications when operating conditions change dramatically, especially when other classifiers only have a fixed set of bad candidates from which to choose. Our approach to distributional anomaly detection is to gather local information using features tailored to the domain, aggregate all such evidence to form a global ii

Robust Anomaly Detection with Applications to Acoustics and Graphs

This paper describes a new algorithm, the NP algorithm, for detecting speech signals of varying quality and gain levels. NP operates in the frequency domain and renders speech/no-speech decisions based on a signal-to-noise ratio (SNR) derived from a sorted power spectrum. In addition to the SNR estimates, a spectral whitening process and an estimate of the variance in the ratio of the signal power to total energy are also used to identify and reject signals that are stationary or nearly stationary The key features of this algorithm are: the detection is based on a single FFT; decisions are independent of the signal gain; the process has 3 dB/octave processing gain from the transform; and frequency domain processing permits the exploitation of the signal structure.

J. Pencak

Papers

The NP speech activity detection algorithm