Tone recognition in Thai has been extensively researched at Chulalongkorn University. Tungthangthum (1998) extracted pitch information from Thai monosyllables using autocorrelation and used HMM for tone classi-fication. An evaluation on small sets of monosyllables achieved 90% accuracy. Based on the analysis by Gandour et al. (1994) described above, Thubthong et al. (2001c)
introduced a novel feature called half-tone modeling, which extended an analysis frame to cover half of the neighboring frames. Application of the proposed feature in a statistical classifier based on a posteriori probabilities yielded approximately a 3% improvement over the baseline model without half-tone modeling. A successive work by Thubthong et al.(2001b) conducted a combination of neural networks with
distinctive features for F0 slope and height for tone recognition.
Outputs from three neural networks, trained by three different sets of distinctive features, were combined for the final decision. The proposed model yielded as much
as a 15.6% error reduction compared to the use of single neural network. Potisuk et al. (1995) conducted an analysis–synthesis approach for tone recognition in continuous speech. The method exploited an extension of Fujisaki’s model to synthesize F0 contours of possible tone sequences of a given speech signal. The resulting contours were then compared to the smoothed and normalized F0 contour of the input speech in a pattern matching module. Overall, 89% of four tone sequences were correctly classified with this method. In 1998, Kongkachandra et al. (1998) proposed a novel feature in harmonic frequency domain for tone recognition. The algorithm was to segment the frequency spectrum into overlapped frames and calculate amplitude difference values between adjacent frames. Classification
of tones was achieved by applying some heuristic rules on the amplitude difference values. The model gave 90% accuracy on a monosyllable test set. All the mentioned
models however, were evaluated on either isolated syllables or hand-segmented syllables from continuous speech. Further research on segmenting speech signals into sequences of syllables or phonemes is required in order to achieve a completely automatic tone recognizer for continuous speech.