engleski

Rapid Prototyping of a Croatian Large Vocabulary Continuous Speech Recognition System

The Croatian language, like many minority languages used by less than 0.1% of the world population, is in need of mature automatic speech recognition (ASR) systems for applications such as transcription of speech recordings, voice control, an aid to impaired people, etc. This paper describes a short-term research and development project aimed to produce an applicable Croatian large vocabulary continuous speech recognition system from scratch. The open-source CMU Sphinx toolkit was our platform choice. For the purpose of acoustic model training, we made a speech training set of several hundred utterances, containing words carefully chosen according to their phonetic properties. Language models were derived from the Croatian large-scale n-gram system, which ensures the system’s applicability. During the project, we succeeded in developing an ASR system able to recognize freely chosen utterances composed of 15, 000 most frequently used Croatian words reasonably well.

automatic speech recognition; continuous speech; large-scale n-gram model; large vocabulary.

nije evidentirano