Introduction
Building speech-to-speech (S2S) translation systems typically requires massive amounts of annotated data. For languages like Thai, which feature complex tonal structures and a non-Roman script, the "cold start" problem is particularly acute. This research presents a prototype Thai-English S2S system designed for medical consultations, demonstrating that functional, high-accuracy tools can be developed even with scarce training data—specifically, just six hours of recorded speech.
Methodology
The system architecture is modular, comprising three primary components: Automatic Speech Recognition (ASR), Machine Translation (MT), and Text-to-Speech (TTS).
Automatic Speech Recognition (ASR)
To overcome data scarcity, the researchers employed a "bootstrapping" technique. Acoustic models from seven diverse languages (including Chinese, German, and Turkish) served as "seeds" to help the system identify Thai phonemes. Furthermore, the Thai script was romanized based on pronunciation. This romanization bypassed the engine's inability to process Thai characters and allowed non-Thai researchers to contribute to grammar development.
Machine Translation (MT) and Interlingua
The translation component utilizes an Interlingua approach known as the Interchange Format (IF). Instead of direct language-to-language mapping, the system translates Thai input into a neutral, intermediate representation of meaning. This architecture facilitates expansion to other languages and enables "paraphrasing," where the system repeats the recognized intent back to the user to verify accuracy before final translation.
Speech Synthesis (TTS)
The Thai voice was constructed using the Festival Speech Synthesis System, employing unit-selection concatenative synthesis. To ensure fluency and handle the tonal nature of the language, each phone was tagged with syllable and tone information.
Experimental Setup
The study utilized a high-quality dataset consisting of 6 hours of Thai text and speech data within the hotel reservation and medical domains. The training set included 34 speakers, while development and evaluation sets each featured 4 speakers. The system vocabulary was constrained to 734 words.
Results
The bootstrapped approach yielded impressive results despite the limited dataset. The Thai speech recognizer achieved a word accuracy of 85.62% for context-independent models on the development set.
As shown in the table above, performance remained robust on the evaluation set, with accuracies ranging from 82.71% to 84.44%. These results indicate that a functional translation prototype can be built with a "reasonable" amount of data (6 hours) while exceeding the 80% accuracy threshold required for practical use.
Discussion and Future Work
While the prototype demonstrates feasibility, several limitations remain. The high accuracy is partly attributed to the closed vocabulary of the target domains; performance in open-ended conversation may vary. Future work will focus on explicit tonal modeling in the acoustic engine, expanding hand-written grammars for broader coverage, and evaluating statistical versus knowledge-based approaches for natural language generation.
Summary updated