Metody neřízené adaptace online rozpoznávání­ řeči

This paper deals with adaptation techniques based on maximum likelihood linear transformations, which are well suited for the task of on-line recognition. When transcriptions are available before the system starts running, we are speaking about supervised adaptation. In unsupervised adaptation the transcriptions have to be computed in the first pass of the recognition process. This is often the case in on-line recognition, where data are gathered continuously. Because the system does not work perfectly it is suitable to assign a certainty factor (CF) to each of the transcriptions. Only data that transcriptions have high CF are used for the adaptation. In the on-line recognition, the adaptation (in the sense of updating transformation formulas) has to be performed iteratively whenever the amount of recognized data reaches the pre-specified level. When small amount of adaptation data is available, it is suitable to involve regression trees to cluster similar model parameters. It is quite useful to adapt both speech and silence parameters. Because speech and silence have very different characteristics, we have separated them into two different clusters. Presented methods have been tested on short term recordings and results have proved the suitability of the proposed approach