ABSTRACT
In this paper, a new calculating method of generalized dimension, based on boxes-coupling principle, is proposed to overcome the edge effects and to improve the capability of the speech endpoint detection which is based on the original calculating method of generalized dimension. This new method has been applied to speech endpoint detection. Firstly, the length of overlapping border was determined, and through calculating the generalized dimension by covering the speech signal with overlapped boxes, three-dimension feature vectors including the box dimension, the information dimension and the correlation dimension were obtained. Secondly, in the light of the relation between feature distance and similarity degree, feature extraction was conducted by use of common distance. Lastly, bi-threshold method was used to classify the speech signals. The results of experiment indicated that, by comparison with the original generalized dimension (OGD) and the spectral entropy (SE) algorithm, the proposed method is more robust and effective for detecting the speech signals which contain different kinds of noise in different signal noise ratio (SNR), especially in low SNR.
Subject(s)
Humans , Artificial Intelligence , Pattern Recognition, Automated , Methods , Signal Processing, Computer-Assisted , Speech , Speech Production Measurement , Methods , Speech Recognition SoftwareABSTRACT
Automatic speaker gender identification based on voice feature is an important task in voice processing and analysis fields. In this paper non-linear parameters such as fractal dimension are applied to be one part of feature space for improving the ability of describing speaker gender feature through conventional linear parameters method. Pitch is picked using lifting scheme, and audio fractal dimension is extracted. Then based on Takens theory, the time delay method is used to reconstruct the phase space of fractal dimension sequence. And fractal dimension complexity is obtained by calculating Approximate Entropy. Three dimension feature vectors, including the pitch, the fractal dimension and the fractal dimension complexity, are applied to speaker gender identification. Experiment results show that through adding non-linear parameters, compared with the linear parameter using one dimension only such as pitch, the proposed method is more accurate and robust, and thus provides a new way for speaker gender identification.