A Study On Robust Speech Endpoint Detection Algorithms In Noisy Environment

While automatic speech recognition (ASR) is applied in adverse acoustic environment, the imparity between the experimental and actual condition degrades the capability of ASR system. Endpoint detection is the first crucial technology, whose accuracy determines the success of the whole SR system to some extent. The detection of the presence of speech embedded in various types of non-speech events and background noise is called endpoint detection, speech detection or voice activity detection(VAD).The need for speech detection also occurs in adaptive speech coding and speech enhancement.Conventional approaches to VAD are put forward according to the silent lab, while new VAD methods in noisy surroundings come out until recent two years. In order to meet the need of ASR's widely use, it is very important to study robust VAD means.In this paper, many VAD algorithms are presented firstly, which include both internal and external methods during the past 10 years. With simulation result and some improvements shown, several VAD features are analyzed too. Then, two new VAD approaches in noisy surroundings are proposed. The first one: a method based on LPCCMCC for endpoint detection of noisy speech signal. It takes Lpc Mel cepstral (LPCCMCC) as feature parameter instead of normal cepstral feature; LPCCMCC considers both auditory and vocal tract character. The experiments show good detection capability with a small SNR can be obtained. The second one: a robust two-stage method based on Chinese phonetic knowledge. Short-time EZQ (energy- zero- quotient) and short-time spectra amplitude (200-1000HZ) are adopted as detection features; adaptive threshold and combination of Chinese phonetic knowledge are also introduced to the detection process. The experiments show better robust capability and high efficiency can be obtained in common noisy environments. At the end of the article, some new study fields within the past two years are come up with and developing perspective of VAD is referred to.