A Hidden Conditional Random Field-Based Approach for Thai Tone Classification

Abstract

In Thai, tonal information is a crucial component for identifying the lexical meaning of a word. Consequently, Thai tone classification can obviously improve performance of Thai speech recognition system. In this article, we therefore reported our study of Thai tone classification. Based on our investigation, most of Thai tone classification studies relied on statistical machine learning approaches, especially the Artificial Neural Network (ANN)-based approach and the Hidden Markov Model (HMM)-based approach. Although both approaches gave reasonable performances, they had some limitations due to their mathematical models. We therefore introduced a novel approach for Thai tone classification using a Hidden Conditional Random Field (HCRF)based approach. In our study, we also investigated tone configurations involving tone features, frequency scaling and normalization techniques in order to fine-tune performances of Thai tone classification. Experiments were conducted in both isolated word scenario and continuous speech scenario. Results showed that the HCRF-based approach with the feature F_dF_aF, ERB-rate scaling and a z-score normalization technique yielded the highest performance and outperformed a baseline using the ANNbased approach, which had been reported as the best for the Thai tone classification, in both scenarios. The best performance of HCRF-based approach provided the error rate reduction of 10.58% and 12.02% for isolated word scenario and continuous speech scenario respectively when comparing with the best result of baselines.

Acoustic-Phonetic Approaches for Improving Segment-Based Speech Recognition for Large Vocabulary Continuous Speech

“…More explicit extraction of Acoustic-Phonetic constraints such as formant tracks, segment duration modeling, and places of articulation could be incorporated to the scoring of the segment graphs. Furthermore, other modeling techniques with higher levels of model complexities such as Conditional Random Fields [30,31] and Deep Neural Networks [32] could be experimented to replace the modeling with Gaussian Mixtures. Evaluating phoneme recognition performances based purely on the acoustic of the speech signals should be a good indicator to the performance of phoneme recognition tasks in which higher-level constraints cannot be used.…”

Acoustic Features for Hidden Conditional Random Fields--Based Thai Tone Classification

“…Although our HCRF-based approach in our previous work [Kertkeidkachorn et al 2014b] outperforms other approaches, especially an ANN-based approach, which had been reported as the best approach for Thai tone classification, our acoustic features are limited to F 0 values and their derivatives.…”

“…Maleerat et al reported that they achieved an accuracy of 91.4% for Thai tone recognition on the isolated-word scenario. Recently, we introduced the novel approach for Thai tone classification on both the isolated-word scenario and the continuous-speech scenario [Kertkeidkachorn et al 2014b]. Instead of using conventional approaches, such as an ANN-based approach or an HMM-based approach, we selected the HCRF-based approach for Thai tone classification because it has three advantages that overcome other approaches: (1) the HCRF-based approach can support sequential data, especially time series of F 0 values and their derivative;…”

“…In Thai, the studies of Thai tone classification Tan et al 2004;Maleerat et al 2009;Tungthangthum 1998], including our previous work [Kertkeidkachorn et al 2014b], mostly focused on acoustic cues limited to F 0 values and their derivatives in the acoustic feature extraction process. Nonetheless, other acoustic cues could also contribute to Thai tone classification.…”

“…In our study [Kertkeidkachorn et al 2014b], we adopted a Hidden Conditional Random Field (HCRF)-based approach for Thai tone classification. There are three advantages for selecting the HCRF-based approach: (1) an HCRF-based classifier is a sequential model supporting sequential data properly; (2) the HCRF-based classifier does not assume the strong independence assumption since feature functions are used to capture relations among observations instead of defining the independence assumption, as reported by Gunawardana et al [2005]; and (3) the HCRF-based classifier contains hidden states, which can capture an intrinsic substructure of acoustic cues that swiftly and thoroughly changes in each syllable.…”