The Development of LOTUS-TRD: A Thai Regional Dialect Speech Corpus

• Main Authors: Thatphithakkul, Sumonmas, Thangthai, Kwanchiva, Chunwijitra, Vataya
• Format: Conference Proceeding
• Language: English
• Subjects: Automatic speech recognition; Data models; dialect ASR; Dictionaries; Error analysis; Licenses; low resource language; Recording; speech corpus; speech recognition; Thai dialect; Translation
• ISSN: 2472-7695
• DOI: 10.1109/O-COCOSDA64382.2024.10800335

Although, Thai dialect speech corpus for Automatic Speech Recognition (ASR) is publicly available [1], to accelerate the efficiency of ASR model is still need more speech hours of Thai dialect data in various domains. We present the 180 hours of the LOTUS-TRD comprised the four Thai regional dialects: Northern Thai (NT), Central Thai (CT), Northeastern Thai or Isan (NET) and Southern Thai (ST) speech accompanied with the Thai text transcription of each dialect pronunciation style. The 560 native speakers from 4 dialects, 19-50 years old, were requested to read a set of 100 sentences from their dialect for recording which were randomly selected from 2,500 sentences per dialect. After that the correspondence between sentences and speaker's speech were verified and corrected to prepare for the pronunciation dictionary construction. The number of word tokens occurring in each dialect is quite resemblance. All of word tokens in the corpus are 1,587,691 tokens. The LOTUS- TRD is released under an open license based on CC-BY-SA 4.0. In addition, we provide a baseline result that demonstrates an average Word Error Rate (WER) of 17.10% on the test set.