A hybrid NN-HMM system for connected digit recognition over telephone in Romanian language
Inge Gavăt, M. Zirra, Valentin Enescu · 2002
The paper presents a system for connected speech recognition with a limited number of words over the telephone line. The system, based on a hybrid structure (a hidden Markov model and an artificial neural net) takes into account the speaker and the telephone equipment variability by incorporating a preprocessing stage that subtracts the personal differences. Experiments performed on 100 speakers, with different system configurations: HMM, HMM+ANN, or with the entire system configuration, revealed an improvement of the error rate from 2% to 0.6% for the isolated digit recognition task and from 12% to 4% for connected strings.