Research Of Speech Endpoint Detection Based On Spectral Entropy

Nowadays, in various comminication systems, speech communication service is of the most important ones. These years, many scholars have focused on how to separate the useful speech from the unuseful noise.Detect the starting point and ending point from a section of speech signal, this procedure is called speech endpoint detection, in this way we can separate the useful speech signal and the useless noise. As a preprocessing technology of speech signal, speech endpoint detection plays an important part in practical applications. For erasing the effects of background noise, the accurate of speech processing can be improved. It's widely used in speech recognition, speech amplification, speech coding, echo cancellation.Through several years of research, there are many speech endpoint detection approaches been proposed which can be divided into two classes. One is based on threshold value, as speech signal and noise signal have different features, we can extract the features of each speech section, and compare the value of the features to the threshold value, then obtain the purpose of speech endpoint detection. The earliest ones include short-time energy, zero-crossing rate and so on.The other one is based on model recognition, in this way we need to estimate the model parameter of each speech signal and noise signal for detecting. The princapal based on threshold value is simple and ease to calculate, so it is widely used. However, if the signal-to-noise ratio (SNR) is too low, the speech signal can be submerged in the noise, and its detecting effects can be worse. And the approaches based on model recognition are far more likely to be used in real-time speech signal system for its complication and huge compution.In this paper, we combined the algorithm of adaptive band-partitioning spectral entropy with adaptive filter.To obtain the purpose of reducing noise first and then detection, compared to other approaches, this one is more robust and can operate effectively under low SNR environment.In the paper, we used matlab to simulate the algorithm,and the result demonstrates its robustness.moreover,we choose TMS320VC5416 as processor and TLV320AIC23 as codec chip to study the hardware system of the speech endpoint detection.