Conclusion

This paper considered the problem of decomposing a signal sequence into multiple signal sources, and proposed a method based on the Ergodic HMM. As an example of this problem, the speaker classification problem was considered, and speaker classification experiments were carried out. The following results were obtained.

1. The initial parameters of the Ergodic HMM are important in determining the segmentation boundaries and the category simultaneously.

2. Among the initial HMM parameters, the symbol output probability is the most important in obtaining good performance.

3. In the speakers classification problem, an excellent classification rate is obtained by using the LPC long cepstrum ( the LPC analysis window length was 341ms in these experiments).

4. The average speaker classification rate is improved by selecting the Ergodic HMM that has high likelihood.