Synchronous Alignment

In speaker verification, the maximum Likelihood between criterion is generally used to verify the claimed identity. This is done using two independent models, i.e. a Client model and a World model. It may be interesting to make both models share the same topology, which represent the phonetic underlying structure, and then to consider two different output distributions corresponding to the Client/World hypotheses. Based on this idea, a decoding algorithm and the corresponding training algorithm were derived. The first experiments show, on a significant telephone database, a small improvement with respect to the reference system, we can conclude that at least synchronous alignment provides equivalent results to the reference system with a reduced complexity decoding algorithm. Other important perspectives can be derived.

Related material