|
--- |
|
license: cc-by-sa-4.0 |
|
language: |
|
- th |
|
metrics: |
|
- cer |
|
- wer |
|
library_name: espnet |
|
pipeline_tag: automatic-speech-recognition |
|
--- |
|
# Model Card for Model ID |
|
|
|
<!-- Provide a quick summary of what the model is/does. --> |
|
|
|
This is the baseline model of Khummuang in [Thai-dialect corpus](https://github.com/SLSCU/thai-dialect-corpus). |
|
|
|
The training recipe was based on wsj recipe in [espnet](https://github.com/espnet/espnet/). |
|
|
|
|
|
|
|
### Model Description |
|
|
|
<!-- Provide a longer summary of what this model is. --> |
|
|
|
This model is a Hybrid CTC/Attention model with pre-trained HuBERT encoder. |
|
|
|
The model was pre-trained on Thai-central, Khummuang, Korat, and Pattani and fine-tuned on Khummuang, Korat, and Pattani. (Experiment 3 in the paper) |
|
|
|
We provide some demo code to do inference with this model architecture on colab [here](https://colab.research.google.com/drive/1stltGdpG9OV-sCl9QgkvEXZV7fGB2Ixe?usp=sharing). |
|
(Code is for Thai-Central. Please select the correct model accordingly.) |
|
|
|
## Evaluation |
|
|
|
<!-- This section describes the evaluation protocols and provides the results. --> |
|
|
|
For evaluation, the metrics are CER and WER. before WER evaluation, transcriptions were re-tokenized using newmm tokenizer in [PyThaiNLP](https://github.com/PyThaiNLP/pythainlp) |
|
|
|
In this reposirity, we also provide the vocabulary for building the newmm tokenizer using this script: |
|
|
|
```python |
|
from pythainlp import Tokenizer |
|
|
|
def get_tokenizer(vocab): |
|
|
|
custom_vocab = set(vocab) |
|
custom_tokenizer = Tokenizer(custom_vocab, engine='newmm') |
|
return custom_tokenizer |
|
|
|
with open(<vocab_path>,'r',encoding='utf-8') as f: |
|
vocab = [] |
|
for line in f.readlines(): |
|
vocab.append(line.strip()) |
|
|
|
custom_tokenizer = get_tokenizer(vocab) |
|
|
|
tokenized_sentence_list = custom_tokenizer.word_tokenize(<your_sentence>) |
|
``` |
|
|
|
The CER and WER results on test set are: |
|
|
|
|Micro CER|Macro CER|Survival CER|E-commerce WER|Micro WER|Macro WER|Survival WER|E-commerce WER| |
|
|---|---|---|---|---|---|---|---| |
|
|5.35|5.65|6.29|5.02|7.53|8.73|11.38|6.09| |
|
<!-- |8.08|11.51|17.39|5.63|12.18|16.65|25.58|7.72| --> |
|
<!-- |18.17|22.38|31.01|13.75|31.74|37.68|50.54|24.82| --> |
|
|
|
## Acknowledgement |
|
|
|
We would like to thank the PMU-C grant (Thai Language Automatic Speech Recognition Interface for Community E-Commerce, C10F630122) |
|
for the support of this research. |
|
We also would like to acknowledge the Apex compute cluster team which provides compute support for this project. |
|
|
|
## Paper |
|
|
|
[Thai Dialect Corpus and Transfer-based Curriculum Learning Investigation for Dialect Automatic Speech Recognition](https://www.isca-speech.org/archive/pdfs/interspeech_2023/suwanbandit23_interspeech.pdf) |
|
``` |
|
@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} |
|
} |
|
``` |
