**Project description:**

The aim of this project was to develop a method which evaluates the
probability that two given speech signals have been spoken by the same
person. The project started from a standard text-dependent speaker
verification method that is based on dynamic time warping and cepstral
distance measurement. This method assumes that the two speech signals
are identically worded. In many practical situations this is not the
case, however. Therefore, the two following problems were
studied within this project:

- How can we automatically and reliably find identically worded segments in two speech signals?
- What type of distance measure has to be used for an optimal discrimination between the two hypotheses same speaker and different speakers?
- How does this method compare to statistical modeling approaches which are normally used for text-independent speaker verification.

**Supported by:** This project was partly supported by
NCCR IM2 (i.e. by the
Swiss National Science Foundation).