Speech Endpoints Detection Method Based On Sub-bands Energy And Pitch Characteristics

The main target of the speech endpoints detection technique is to getrid of non-voice signal from the streams of sampling signal, and to makesure the starting and ending points of the speeches. This technique plays akey role in detecting voice and is an important part of the speechrecognition system. An effective and precise detection of speechendpoints can not only decrease the processing data, shorten theprocessing time, but can also get rid of the interferences to thepost-processing of speech, which is originated from the noisy and silentspeech, and thus achieve the goal of improving the speech recognitionperformance. Now under the quiet circumstance, many speech endpointsdetection algorithms have good detection performance, but in thecondition with noisy background, the detection performances of thosealgorithms are usually not satisfied. In real life, most of the speech signalis corrupted by different background noise, hence, the research of speechendpoints detection under the noisy environments very important inimproving the performance of speech recognition for the practicalapplication.At first in this thesis, the characteristic of speech signal is introduced,and the speech signal analysis methods in time and frequency domain arebriefly overview then. Secondly, several common methods of speechendpoint detection, such as the method based on short-time averageenergy，the method of dual-threshold detection based on the combinationof short-time energy and short-time zero crossing rate features, themethod based on cepstrums and etc, are briefly illustrated. Theexperiments show that performances of these methods are ideal in thequiet environment, but it become poor in the noisy environment. In orderto overcome the traditional shortcomings in this regards, a speech endpoints detection algorithm based on sub-bands and pitch features isproposed in this thesis, it can make a speech and non-speech decision ofsignal according to the the pitch and its harmonic values, and it can setthe sub-band decision threshold dynamically according to the sub-bandbackground noise, and the decisions make by every sub-band energy ismerged together to make a final decision of speech endpoints. In the endof the thesis, the proposed speech endpoints detection algorithm isapplied to a speech recognition system to test the effectiveness of thealgorithm. The experiment results show that, compared to thetraditional endpoints detection algorithms, the proposed algorithm in thisthesis has better speech detection performance.