engleski

Croatian Telephone Speech Recognition

In the paper we present speech recognition experiments of Croatian telephone speech. Recognition of spontaneous and planned speech in the noisy telephone environment is described. The speech recognition system was trained and tested on almost 6 hours of telephone speech material. The telephone speech was recorded daily from broadcast weather reports spoken by meteorologist on the national radio by telephone. The recorded telephone speech is rich in background noises and includes over 1800 different words spoken by 12 male and female speakers. The paper describes the acoustical model of the recognition system, which is based on context independent and context dependant hidden Markov models. The acoustic modelling procedure, based on Croatian phonetic rules is presented. The language model of the speech recognition system is described and speech recognition results are presented.

acoustic modeling; context dependant phone models; speech recognition

nije evidentirano