Author:
Un-Gian Iunn, Kiat-Gak Lau, Hong-Giau Tan-Tenn, Sheng-An Lee, and Cheng-Yan Kao

Abstract:
A sizable corpus of Taiwanese text in Latin script has been accumulated over the past two hundred or so years. However, due to the special status of Taiwan, few people can read these materials at present. It is regrettable that the utilization of these plentiful materials is very low. This paper addresses problems raised in the Taiwanese Southern-Min tone sandhi system by describing a set of computational rules to approximate this system, as well as the results obtained from its implementation. Using the romanized Taiwanese Southern-Min text as source, we take the sentence as the unit, translate every word into Chinese via an online Taiwanese-Chinese dictionary (OTCD), and obtain the part-of-speech (POS) information from the Chinese Electronic Dictionary (CED) made by the Chinese Knowledge and Information Processing (CKIP) group of Academia Sinica. By using the POS data and tone sandhi rules based on linguistics, we then tag each syllable with its post-sandhi tone marker. Finally, we implement a Taiwanese Southern-Min tone sandhi processing system which takes a romanized sentence as an input and then outputs the tone markers. Our system achieves 97.39% and 88.98% accuracy rates with training and test data, respectively. Finally, we analyze the factors influencing error for the purpose of future improvement.

Keyword: Taiwanese Southern-Min, Written Taiwanese, Tone Sandhi System, Taiwanese Romanization

Author:
Hung-Yan Gu, Yan-Zuo Zhou, and Huang-Liang Liau

Abstract:
In this paper, a framework for integrated synthesis of Mandarin, Min-nan, and Hakka speech is proposed. To show its feasibility, an initial integrated system has been built as well. Through integration, a model only trained with Min-nan sentences is used to generate pitch-contours for all three languages, same rules are used to generate syllable duration and amplitude values, and the same program module implementing the method, TIPW, is used to synthesize the three languages’ speech waveforms. Also, in this system, each syllable of a language has just one recorded signal waveform, i.e. no chance of unit selection. Under such a restricted situation, the synthetic speech signals still have noticeable naturalness level and signal clarity.

Keyword:
Speech Synthesis, Pitch Contour Model, TIPW, Time Axis Warping.

Author:
Wei-Chih Kuo, Chen-Chung Ho, Xiang-Rui Zhong, Zhen-Feng Liang, Hsiu-Min Yu, Yih-Ru Wang, and Sin-Horng Chen

Abstract:
In this paper, three studies of Min-Nan speech processing are presented. The first study concerns the implementation of a high-performance Min-Nan TTS system. On the basis of the waveform templates of 877 base-syllables used as basic synthesis units and through the application of the RNN-based prosody generation method and the PSOLA algorithm for prosody modification, this Min-Nan TTS system can convert texts, represented in both Han-Luo (漢羅) and Chinese logographic writing systems, into natural Min-Nan speech. An informal, subjective listening test confirms that the system performs well and the synthetic speech sounds natural for well-tokenized Min-Nan texts and for automatically tokenized Chinese logographic texts. The second investigation concerns the realization of a Min-Nan speech recognizer. It adopts the initial-final-based HMM approach with a simple base-syllable bigram language model. A base-syllable recognition rate of 65.1% has been achieved. Finally, a model-based tone labeling method is presented. This method adopts a statistical model to eliminate the affections of all factors other than tone on the syllable pitch contour for automatic tone labeling. Experimental results confirm that this method outperforms the conventional VQ-based approach.

Keyword:
Min-Nan Text-to-Speech System, Speech Recognition, Model-Based Tone Labeling

Author:
Jane S. Tsay

Abstract:
Taiwanese Child Language Corpus (TAICORP) is a corpus based on spontaneous conversations between young children and their adult caretakers in Minnan (Taiwan Southern Min) speaking families in Chiayi County, Taiwan. This corpus is special in several ways: (1) It is a Minnan corpus; (2) It is a speech-based corpus; (3) It is a corpus of a language that does not yet have a conventionalized orthography; (4) It is a collection of longitudinal child language data; (5) It is one of the largest child corpora in the world with about two million syllables in 497,426 lines (utterances) based on about 330 hours of recordings. Regarding the format, TAICORP adopted the Child Language Data Exchange System (CHILDES) [MacWhinney and Snow 1985; MacWhinney 1995] for transcribing and coding the recordings into machine-readable text. The goals of this paper are to introduce the construction of this speech-based corpus and at the same time to discuss some problems and challenges encountered. The development of an automatic word segmentation program with a spell-checker is also discussed. Finally, some findings in syllable distribution are reported.

Keywords:
Minnan, Taiwan Southern Min, Taiwanese, Speech Corpus, Child Language, CHILDES, Automatic Word Segmentation

Author:
Jiang-Chun Chen, Jyh-Shing Roger Jang, and Te-Lu Tsai

Abstract:
This paper presents the algorithms used in a prototypical software system for automatic pronunciation assessment of Mandarin Chinese. The system uses forced alignment of HMM (Hidden Markov Models) to identify each syllable and the corresponding log probability for phoneme assessment, through a ranking-based confidence measure. The pitch vector of each syllable is then sent to a GMM (Gaussian Mixture Model) for tone recognition and assessment. We also compute the similarity of scores for intensity and rhythm between the target and test utterances. All four scores for phoneme, tone, intensity, and rhythm are parametric functions with certain free parameters. The overall scoring function was then formulated as a linear combination of these four scoring functions of phoneme, tone, intensity, and rhythm. Since there are both linear and nonlinear parameters involved in the overall scoring function, we employ the downhill Simplex search to fine-tune these parameters in order to approximate the scoring results obtained from a human expert. The experimental results demonstrate that the system can give consistent scores that are close to those of a human’s subjective evaluation.

Keyword:
CAPT, CALL, Speech Recognition, Tone Recognition, Speech Assessment, GMM, Mandarin Chinese, Downhill Simplex Method, Phoneme, Intensity, Rhythm, Forced Alignment
　

Abstract:
Coreference resolution is the process of determining the entity that noun phrases refer to. A great deal of research has been done on this task in English, using approaches ranging from those based on linguistics to those based on machine learning. In Chinese, however, much less work has been done in this area. One reason for this is the lack of resources for Chinese natural language processing. This paper presents a knowledge-based, unsupervised clustering algorithm for Chinese coreference resolution that maximizes performance using freely and easily available resources. Experiments to demonstrate the efficacy of such an approach are performed on two data sets: TDT3 and ACE05, and the ACE value coreference resolution results achieved through our approach are 52.5% and 55.2% respectively. An oracle experiment using gold standard noun phrases achieved even more impressive results of 77.0% and 76.4%. To analyze the causes of errors, this paper also looks into false alarms and misses in documents.

Keyword:
Coreference Resolution, Modified K-means Clustering, Stacked Transformation-based Learning, Unsupervised Learning