A speech recognition algorithm based on the features of Croatian language

The paper examines an acoustic-phonetic, approach to speech recognition, with elements of artificial intelligence. The vowels of the Croatian language are assumed as examples of speaking segments. Frequency transformation has been realized passing the speech segments through the corresponding selective filter banks. It has been shown that on the outputs of this filter bank the significantly comparable speaker independent signals can be obtained when the filter resonant frequencies are multiplied by the appropriate speaker dependent number. The block scheme of the speech recognition device, with SC filters in the role of the filter bank, the neural network, and the signal processor, is proposed and discussed at the end.