Automatic recognition of oral vowels in tone language: Experiments with fuzzy logic and neural network models.
The standard Yoruba (SY) language is used for our case study.
Automatic recognition of tone language speech is a complex problem in that it involves two parallel recognition tasks. A recognition system to accomplish this task must be able to simultaneously recognise tone and phone Components in the acoustic signal. The acoustic cue for the tones is the fundamental frequency (F0) while the first and second formant (F1 and F2) frequencies are the acoustic cues for the phones. In this study, we experiment with two soft-computing techniques, namely: artificial neural network (ANN) and fuzzy logic (FL) in the recognition of oral vowels in tone language. The standard Yoruba (SY) language is used for our case study.The ANN and FL speech recognition systems were developed using MatLab. The result showed that the ANN based model performed better on the training data while the FL based model performed better on the test set. This implies that the ANN system was able to interpolate or approximate the data more accurately whereas the FL system is better at extrapolating from the data. In addition, it was observed that the ANN system required larger amount of data for it is development whereas the FL system development requires some expert's knowledge. In conclusion, the FL based system seems to be the better approach for developing practical automatic speech recognition (ASR) system for languages such as SY where the language resources are limited.