Chai Wutiwiwatchai

BEST 2009 : Thai word segmentation software contest

This is a non-technical paper describing how and why we organized BEST 2009, the first contest in the series of “Benchmark for Enhancing the Standard of Thai language processing”, which is expected to help accelerate the progress of the Natural Language Processing technology in Thailand by assembling 3 essential components: common standards, resources and researchers. The BEST 2009 : Thai Word Segmentation Software Contest is the first shared task on Thai NLP that exercised this assemblage and aimed to find the best algorithms that could correctly divide Thai non-segmented script into words according to the guidelines previously prepared by experts from several research institutes and universities. Thai word-segmented corpora of 5 million words have been developed as a training set, another 600K as a test set. The evaluation procedure and protocol have been designed. The process and the results of the contest are reported.

A study of VoIP quality evaluation: User perception of voice quality from G.729, G.711 and G.722

This paper presents new evidence about user perception of VoIP quality that is inconsistent with the general understanding of three codecs know as G.729, G.711 and G.722. The focus of the study is aimed at VoIP quality evaluation by Thai users that use the Thai language which is tonal. This study was conducted by using conversation-opinion tests. The results, called MOS-CQS, were then analyzed carefully. After the study and analysis, it has been found that the perception of subjects, who were Thai native speakers, to G.729, G.711 and G.722 is not significantly different.

LOTUS-BN: A Thai broadcast news corpus and its research applications

This paper describes the design and construction of the LOTUS-BN corpus, a Thai television broadcast news corpus. In addition to audio recordings and their transcription, this corpus also includes a detailed annotation of many interesting characteristics of broadcast news data such as acoustic condition, overlapping speech, news topic and named entity. The LOTUS-BN is still an ongoing project with the goal of collecting 100 hours of speech. We report initial statistics analyzed from 60 hours of speech which show that the LOTUS-BN corpus has a rich vocabulary of approximately 26,000 words with one third of them are named entities. Thus, this corpus is a good resource for developing an LVCSR system and investigating on named entity detection and recognition in addition to broadcast news related applications. Research applications on these topics are also discussed.

The Asian network-based speech-to-speech translation system

This paper outlines the first Asian network-based speech-to-speech translation system developed by the Asian Speech Translation Advanced Research (A-STAR) consortium. The system was designed to translate common spoken utterances of travel conversations from a certain source language into multiple target languages in order to facilitate multiparty travel conversations between people speaking different Asian languages. Each A-STAR member contributes one or more of the following spoken language technologies: automatic speech recognition, machine translation, and text-to-speech through Web servers. Currently, the system has successfully covered 9 languages— namely, 8 Asian languages (Hindi, Indonesian, Japanese, Korean, Malay, Thai, Vietnamese, Chinese) and additionally, the English language. The systems domain covers about 20,000 travel expressions, including proper nouns that are names of famous places or attractions in Asian countries. In this paper, we discuss the difficulties involved in connecting various different spoken language translation systems through Web servers. We also present speech-translation results on the first A-STAR demo experiments carried out in July 2009.

The state of the art in Thai language processing

This paper reviews the current state of technology and research progress in the Thai language processing. It resumes the characteristics of the Thai language and the approaches to overcome the difficulties in each processing task.

Text-dependent speaker identification using neural network on distinctive Thai tone marks

Presents a neural network based text-dependent speaker identification system for Thai language. Linear prediction coefficients (LPC) are extracted from speech signal and formed feature vectors. These features are fed into a multilayer perceptron (MLP) neural network with backpropagation learning algorithm for training and identification processes. Five Thai tone marks are considered very closely in choosing the sentences in order to achieve the best speaker identification accuracy. Five speaking texts with each Thai tone and a mixed tone text are comparatively experimented. Average identification rate on 9 speakers achieves above 95% when using mixed tone text, and poor results occur with middle and low tone texts, which usually cause vagueness or unclear voices.

Text-dependent speaker identification using LPC and DTW for Thai language

This paper proposes a text-dependent speaker identification system applied to Thai language. Isolated digits 0-9 and their concatenations are used for speaking text. Linear prediction coefficients (LPC) are extracted and formed as feature vectors represented each speech signal. Dynamic time warping (DTW) is used to measure distances between referenced and evaluated vectors. These distances, indicating nearness of unknown vectors to references, incorporated with the K-nearest neighbor (KNN) decision technique are used to decide who possesses those unknown vectors. The experimental results have shown that the best identification rate for a single digit is 95.83% and the highest rate for concatenated digits of top-3, top-5, and top-7 are 98.75%, 100%, and 99.20%, respectively.

An Audio Watermarking Scheme Based on Singular-Spectrum Analysis

This paper proposes a blind audio watermarking scheme based on singular-spectrum analysis (SSA) which relates to several techniques based on singular value decomposition (SVD). SSA is used to decompose a signal into several additive oscillatory components where each component represents a simple oscillatory mode. The proposed scheme embedded a watermark into a host signal by modifying scaling factors of certain components of the signal. Test results show that the proposed scheme satisfies imperceptibility criterion suggested by IHC with the average ODG of 0.18. It is robust against many attacks, such as MP3 and MP4 compression, band-pass filtering, and re-sampling. This paper does not only propose a new watermarking scheme, it also discusses the singular value and reveals its meaning, which has been deployed and played an important role in all SVD-based schemes.

VoIP Quality of Experience: A proposed subjective MOS estimation model based-on Thai users

To aid authenticity, accuracy and reliability of subjective measurement of VoIP quality measurements, this paper proposes a model that has been created from the subjective MOS, instead of using existing objective measurement methods for VoIP quality measurement. The proposed model of VoIP quality measurement method is based-on Thai users. The data has been gathered from conversation-opinion tests with 400 Thai subjects, referring to packet loss and packet delay effects. This model is called the Thai subjective VoIP Quality Evaluation model (ThaiVQE). It has been evaluated by testing with a test set of subjective MOS, then the results have been compared with the E-model results. Based-on Thai subjects, the evaluation result shows that ThaiVQE can provide better accuracy and reliability than the E-model with improvement of over 20%.

Thai ASR development for network-based speech translation

A network-based multilingual speech translation service under the Universal Speech Translation Advanced Research (U-STAR) consortium requires a well-tuned Thai automatic speech recognition (ASR) service. This paper summarizes the development of the service by utilizing both Thai read-speech and telephone speech (LOTUS-CELL 2.0) corpora. Tuning is performed regarding different sets of acoustic unit and training data. An evaluation shows that the recognition accuracy of ASR working over data channels can be improved by using the LOTUS-CELL 2.0 corpus although the corpus was constructed via voice channels. The problem of Named-entity (NE) words often found in the working domain is obvious and leads to an urgent future work.