In this thesis work, we design and implement a robust isolated word recognition system which can be used in the presence of car-noise. The system uses word-level hidden Markov models (HMMs) for designated command vocabularies to control various functions of a car.
In oder to improve the performance of the speech recognition system, we do the following steps. First, we propose a new method for end-point detection by using linear prediction error energy. Second, we examine six feature parameters to select the most robust feature parameter in the presence of car-noise. Third, we allocate the number of states in HMMs for each isolated word proportional to the average number of analysis frames in the word. Finally, we propose a car-noise dependent vector quantization algorithm to adapt the variant car-noise environment.
The performed word recognition system is tested using 31 Korean isolated words to control function of the accessories in a car. Test results show that the recognition performance of the proposed system is considerably improved as compared with those already established recognition systems.
