---
language:
- th
tags:
- automatic-speech-recognition
license: apache-2.0
datasets:
- common_voice
metrics:
- wer
- cer
---

# Thai Wav2Vec2 with CommonVoice V8 (newmm tokenizer)

This model trained with CommonVoice V8 dataset by increase data from CommonVoice V7 dataset that It was use in [airesearch/wav2vec2-large-xlsr-53-th](https://huggingface.co/airesearch/wav2vec2-large-xlsr-53-th). It was finetune [wav2vec2-large-xlsr-53](https://huggingface.co/facebook/wav2vec2-large-xlsr-53).

GitHub: [https://github.com/wannaphong/thai-wav2vec2-cv-v8](https://github.com/wannaphong/thai-wav2vec2-cv-v8)

## Datasets

It is increase new data from The Common Voice V8 dataset to Common Voice V7 dataset or remove all data in Common Voice V7 dataset before split Common Voice V8 then add CommonVoice V7 dataset back to dataset.

It use [ekapolc/Thai_commonvoice_split](https://github.com/ekapolc/Thai_commonvoice_split) script for split Common Voice dataset.

You can read more at [wannaphong/thai_commonvoice_dataset](https://github.com/wannaphong/thai_commonvoice_dataset)

## Models

This model was finetune [wav2vec2-large-xlsr-53](https://huggingface.co/facebook/wav2vec2-large-xlsr-53) model with Thai Common Voice V8 dataset and It use pre-tokenize with `pythainlp.tokenize.word_tokenize`.

## Training

I used many code from [vistec-AI/wav2vec2-large-xlsr-53-th](https://github.com/vistec-AI/wav2vec2-large-xlsr-53-th) and I fixed bug training code in [vistec-AI/wav2vec2-large-xlsr-53-th#2](https://github.com/vistec-AI/wav2vec2-large-xlsr-53-th/pull/2)

## Evaluation

**Test with CommonVoice V8 Testset**

| Model                 | WER by newmm (%) | WER by deepcut (%) | CER      | URL                                                         |
|-----------------------|------------------|--------------------|----------|-------------------------------------------------------------|
| wav2vec2 with deepcut | 16.354521        | 11.424476          | 3.684060 | https://github.com/wannaphong/th-cv-v8-wav2vev2-deepcut     |
| wav2vec2 with newmm   | 16.698299        | 11.436941          | 3.737407 | https://github.com/wannaphong/thai-wav2vec2-cv-v8           |
| CV v7                 | 17.414503        | 11.923089          | 3.854153 | https://huggingface.co/airesearch/wav2vec2-large-xlsr-53-th |

**Test with CommonVoice V7 Testset (same test by CV V7)**

| Model                 | WER by newmm (%) | WER by deepcut (%) | CER      | URL                                                         |
|-----------------------|------------------|--------------------|----------|-------------------------------------------------------------|
| wav2vec2 with deepcut | 12.776381        | 8.773006           | 2.628882 | https://github.com/wannaphong/th-cv-v8-wav2vev2-deepcut     |
| wav2vec2 with newmm   | 12.750596        | 8.672616           | 2.623341 | https://github.com/wannaphong/thai-wav2vec2-cv-v8           |
| CV v7                 | 13.936698        | 9.347462           | 2.804787 | https://huggingface.co/airesearch/wav2vec2-large-xlsr-53-th |

This is use same testset from [https://huggingface.co/airesearch/wav2vec2-large-xlsr-53-th](https://huggingface.co/airesearch/wav2vec2-large-xlsr-53-th).

source code benchmark: https://github.com/wannaphong/thai-asr-benchmark/tree/main/commonvoice


## Files
- `0-download-unzip.ipynb` - notebook for download and unzip CommonVoice V8
- `1-convert-mp3-wav.ipynb` - notebook for convert mp3 files to wav files
- `1-preprocessing-thai-cv-v8-wav2vev2.ipynb` - notebook for preprocessing CommonVoice V8 (old file)
- `2-gen-val-json.py` - python file for get manifest in nvidia meno asr
- `2-preprocessing-thai-cv-v8-wav2vev2.ipynb` - notebook for preprocessing CommonVoice V8
- `4-gen-manifest.ipynb` - notebook for get manifest in nvidia meno asr
- `build-lm.ipynb` - notebook for build ASR LM
- `test-ai4thai.ipynb` - notebook for test AI For Thai.
- `test-google.ipynb` - notebook for test Google ASR.
- `test-v7.ipynb` - notebook for test [vistec-AI/wav2vec2-large-xlsr-53-th](https://github.com/vistec-AI/wav2vec2-large-xlsr-53-th) model.
- `test-wav2vec2-lm.ipynb` - notebook for test our model with LM.
- `test-wav2vec2.ipynb` - notebook for test our model without LM.
- `train-wav2vec2.py` - python file for training model.

**Links:**
- GitHub Dataset: [https://github.com/wannaphong/thai_commonvoice_dataset](https://github.com/wannaphong/thai_commonvoice_dataset)