Speech Recognition for Bosnian/Croatian/Serbian

Project Type: Master/Diploma Thesis
Student: Friedl Alexander
Mentor: Gernot Kubin

 

 In the present work an automatic, speaker-independent speech recognition system based on Hidden-Markov-models for the south Slavic languages Bosnian, Croatian and Serbian is built. The system is trained and tested with an already existing database of audio files. With respect to these results a new database with files containing human speech of native speakers of the treated languages is realized. More than four and a half hours of speech from 25 speakers are collected. In four predefined test scenarios the speech recognition system is evaluated depending on database size, phoneme model and phoneme combination. In the case of connected word recognition high correctness rates of about 85 percent are achieved. The results of isolated word recognition range between 97 and 100 percent and reach therefore full correctness in many cases.