The application of long-term analysis of the zero-crossing of a speech signal in automatic speaker identification

Downloads

Authors

  • Czesław BASZTURA Institute of Telecommunication and Acoustics, Technical University, Poland
  • Wojciech MAJEWSKI Institute of Telecommunication and Acoustics, Technical University, Poland

Abstract

This paper investigates the possibility of using long-term analysis of the zero-crossings of a speech signal for speaker identification. The applied method of identification is based on an analysis of the signal in measuring windows of a duration which should ensure the stationarity of the statical distributions of the time intervals between successive zero-crossings, in 16 pre-set time channels. An objective method of defining the minimum length of the measuring window for a selected set of parameters is presented. It is based on the stationarity test and the ergodic theorem for stochastic processes, as is the transformation of the speech signal mentioned above. An experiment in speaker identification for 10 speakers with 10 repetitions for each speaker has been performed. The results obtained, in well exceeding 90 % correct identification for 30- and 40 second samples of the speech signal, have confirmed the practicality of the method of zero-crossing analysis for speaker identification.

References

[1] J. S. BENDAT, A.G. PIERSOL, Random data: Analysis and measurement procedures, Wiley, 1971.

[2] R. GUBRYNOWICZ, The zero-crossing method in speech signal analysis, and the automatic identification of a limited set of words, Reports IPPT PAN, No 37/1974 [in Polish].

[3] K-P. LI, G.W. HUGHES, Talker differences as they appear in correlation matrices of continuous speech spectra, Journ.Acoust.Soc.Am., 55, 4, 833-837 (1974).

[4] W. MAJEWSKI, H. HOLLIEN, Euclidean distance between long-term speech spectra as a criterion for speaker identification, Proc. Speech Communication Seminar, Stockholm 1974, Vol. 3, pp. 303-310.