The endpoint detection is a necessary step of speech analysis, speech synthesis and speech recognition. The accuracy of the endpoint detection in quiet experiment environment was inspiring, but when the system in the practical environment, the performance was decreased significantly.In this paper, based on the summarization and comparison the methods, the box-counting dimension of the fractal theory was used in the endpoint detection. The classical threshold method was improved, used the endpoint detection algorithm based on the box-counting dimensions and short-time energy-zero ratio. The box-counting dimension was used as the main judgment, the judging threshold setting was used the self-adaptive mechanism.After the matlab experimental verification, found that this algorithm had high accuracy and good robustness in the white noise and pink noise environment but not very satisfied in the volvo and babble noise environment. |