An Empirical Study for Constructing Thai Tone Models

Tone modelling is a critical compo-nent for speech recognition in Thai language. In this paper, we pre-sented an empirical study examining which configurations with respect to tone features, frequency scale, nor-malization technique, and tone crit-ical segment should be used for tone recognition. We built three contin-uous speech corpora with different complexities. A feedforward neu-ral network was employed to evalu-ate these configurations on all cor-pora. From the experimental results, we conclude that the tone features, containing five F0’s and their slopes at five different time points, yields the highest recognition rate, ERB-rate scale surpasses semi-tone and hertz scales, z-score normalization significantly exceeds mean normaliza-tion, and rhyme units provide bet-ter performance than syllable units. Based on these results, we have pro-posed a Thai tone modelling. The modelling provides 92.86%, 90.05%, 84.07%, and 82.48 % accuracies fo