Multilingual Italian – Lithuanian Small Vocabulary Speech Recognition via Selection of Phonetic Transcriptions

Abstract

The amount of resources required to incorporate a new language into an ASR system is extensive and it is usually hard to provide more than one language per system. Naturally some research groups started developing the multilingual speech recognition principles. Opposite to most of the other researchers, in this work the original recognizer was not modified as we chose to use the phonetic “language” of the recognizer – the “transcriptions”, allowing the adaption of the original ASR engine to any language if enough acoustic resources are available. The results of our experiments proved that it is quite reasonable to use these methods to provide the multilingual speech interface for the user. It was established that the recognition accuracy can be acceptable for small vocabulary grammars (~93% for Lithuanian, and ~98% for Italian), satisfying the requirements of an IVR type of application where different small vocabulary grammars with a limited number of multilingual transcriptions can be loaded for each of the different stage of the HCI voice dialog, allowing the modeling of multilingual regional self-services.

DOI: http://dx.doi.org/10.5755/j01.eee.121.5.1145