Conclusion

A finer optimization should be obtained by combining the stamp value and using an HMM model for smoother estimates.

Our approach shows that this type of coding contains useful information for distinguishing speaker specificities, and we now need to integrate it into a tracking model that should provide complementary information to conventional models. (e.g. Gaussian model with cepstral parameters): SpkDet Mistral module .

The major difference with our qualitative approach lies in a representation for each speaker in a space of very small integers (the identity of 50 speakers is (partially) encoded by a set of QTF vectors each on [1:13] ¹⁵ ).