Model Card for Model ID

This is the baseline model of Thai-central in Thai-dialect corpus.

The training recipe was based on wsj recipe in espnet.

Model Description

This model is a Hybrid CTC/Attention model with pre-trained HuBERT as the encoder.

This model was trained on Thai-central to be used as a supervised pre-trained model in order to be used for finetuning to other Thai dialects. (Experiment 2 in the paper).

We provide some demo code to do inference with this model on colab here. (Please note that you cannot inference >4 seconds of audio with free Google colab)

Evaluation

For evaluation, the metrics are CER and WER. Before WER evaluation, transcriptions were re-tokenized using newmm tokenizer in PyThaiNLP

from pythainlp import word_tokenize

tokenized_sentence_list = word_tokenize(<your_sentence>)

The CER and WER results on the test set are:

CER = 2.0

WER = 6.9

Paper

Thai Dialect Corpus and Transfer-based Curriculum Learning Investigation for Dialect Automatic Speech Recognition

@inproceedings{suwanbandit23_interspeech,
  author={Artit Suwanbandit and Burin Naowarat and Orathai Sangpetch and Ekapol Chuangsuwanich},
  title={{Thai Dialect Corpus and Transfer-based Curriculum Learning Investigation for Dialect Automatic Speech Recognition}},
  year=2023,
  booktitle={Proc. INTERSPEECH 2023},
  pages={4069--4073},
  doi={10.21437/Interspeech.2023-1828}
}