Oscar Saz research description

My current research focuses on the construction of systems for the automatic recognition of BBC data. We are currently working in the 3 parts that define the system, this is, the preparation of the lexicon, the training of acoustic models and the creation of language models.

 

Regarding the work with the lexicon, we are working with the Combilex dictionary and the Sheffield background dictionary (UNISYN-based) towards understanding how to reduce the number of errors due to inaccurate transcriptions and to minimize the problems due to the presence of Out-of-Vocabulary words. In acoustic modeling, in acoustic modeling we are building models with in-domain BBC data that can reliable recognize and transcribe this challenging data. Finally, in language modeling, we will investigate different approaches and models and how they help improve recognition accuracy.

 

In the future, we want to be able to streamline the processing of new and more diverse BBC data, allowing for a prompt and quick recognition and transcription of this data, disregarding its genre, original media or domain. For this, we will use the metadata available in the BBC data to produce canonical models that will be able to deal with all these different conditions in a robust way.