Transcription for Thai Traditional Medicine Texts with Wav2Vec2

Recently, Automatic Speech Recognition (ASR) technology is widely used for communication convenience in converting speech to text. Initial surveys reveal that existing studies on ASR, both in Thai and foreign languages, focus on contemporary language communication, lacking support for specialized vocabulary. Thai Traditional Medicine (TTM) contain ancient medical terminology recorded in various media such as palm leaves, parchment, notebooks, and inscriptions. Centuries of textual preservation have led to text degradation. Given the suitability of ASR study to accommodate medical terminology and ancient languages, it is proposed to apply it to convert original Thai Traditional Medicine texts into contemporary language, aiding experts in reading and interpreting them. ASR involves converting speech signals into text. In this paper, we propose an end-to-end speech recognition approach for Thai Traditional Medical texts leveraging self-supervised learning (SSL) to overcome the limitations of limited labeled data. We utilize a transformer-based model pre-trained (XLSR-Wav2Vec) on unlabeled TTM texts speech data using SSL techniques and fine-tune it with a small dataset of labeled TTM texts audio-text pairs. Due to the absence of a Thai Traditional Medicine vocabulary repository, a dataset comprising text and speech is developed, consisting of 1,800 training sentences and 200 test sentences randomly selected from recordings of 10 Thai Traditional Medicine experts. Offline evaluation methods are employed, resulting in improved accuracy rates. Future efforts will focus on expanding the speech dataset to enhance recognition accuracy further. This study lays the groundwork for efficiently extracting knowledge from Thai Traditional Medicine prescriptions, ensuring the preservation of this valuable heritage for future generations.