Vowel Classification of Children's Speech Using Fundamental and Formant Frequencies

Automatic speech recognition is one of the most revolutionary innovations in telecommunication and speech therapy systems. Vowel classification, especially of children's speech, is a relatively unexplored and challenging dimension of speech recognition and analysis research. In this study, vowel classification of children was investigated for sustained Malay vowels which were produced by Malay children between 7 to 12 years of age, all of whom were categorized into 12 different age and gender groups. Each of the subjects was asked to pronounce all 6 sustained Malay vowels. After a thorough discrimination test, fundamental and formant frequencies (F0- F3) were extracted. Different combinations of fundamental and formant frequencies were examined to find the best combination by applying the Euclidean minimum distance formula on normalized and non-normalized data. Our results showed around 17% improvement when using normalized data. Moreover, different combinations of fundamental and formant frequencies were examined on vowel classification by applying MLP. Based on the results, the best rate of correct classification was 87.50% when using the Euclidean method and 85.79% when using the MLP (Multi-layer Perceptron) method with combinations of fundamental and first three formants frequencies (F0, F1, F2 & F3) in both methods. An additional observation made was the improvement in the trend of accuracy with the increase in the age of the children.