A computational classification of Thai lexical tones

Abstract

The leveling of contour tones in Thai, uttered in a continuous context, has served as a natural point of difficulty for tone recognition experiments. Two tone recognition experiments presented here both include five lexical Thai tones (high, mid, low, rising, and falling) as abstract Bayesian models incorporated into a multi-model Hidden Markov Model. The HMM was developed using Thai natural language utterances to test its performance in correctly identifying Thai lexical tone categories. All utterances used for testing and training were produced in a laboratory setting. Utterances for the first experiment were produced in a citation context, and utterances for the second experiment were produced in a continuous context. The results of the two experiments were compared to test if the context had a significant effect on correctly identifying tone category. Findings showed the context of the utterance had a significant effect on the HMM’s ability to correctly identify tone category, F(1,48) = 5.82, p = 0.020. The identification of the correct tone category for citation utterances showed a significant increase in performance than for the correct identification of tone category for continuous utterances, t(48) = 5.38, p < 0.001.

ATLAS (Automated Tone Level Annotation System): A tonologist's and documentarian's toolkit

“…A common method is based on the Hidden Markov Model (HMM), which is trained on manually annotated data to identify tonal categories. HMMs have been implemented in languages such as Mandarin [4][5], Thai [6], and Cantonese [7]. Other techniques used to address this question include neural networks, which may require less training data and typically aim to increase the speed of transcription for languages with known tone systems [8].…”