Automatic Development of ASR System for an Under-Resourced Language
Radek Šafařík, Lukáš Matějů · 2018
This paper deals with an automatic development of language specific parts (such as vocabulary, language model (LM) and acoustic model (AM)) for Automatic Speech Recognition system. It describes our automatic approach and whole process of adapting of ASR system for Latvian language. We have already developed an approach consisted of methods and tools for adapting our ASR system. The approach utilizes only freely available audio and text data which can be found on the Internet. Vocabulary and LM are automatically created from text downloaded from news websites. AM is created using cross-lingual bootstrapping and lightly supervised re-training process from Latvian TV and radio broadcast data and from archive of Latvian parliament sessions. We showed that it is possible to obtain more than 100 hours of automatically annotated speech in only three months and create system that can achieve performance with WER values in range from 15.3 to 27.8 % with minimum human effort.