asr_ewondo
This is the model for Wav2Vec2-Large-XLSR-Ewondo, a fine-tuned facebook/wav2vec2-large-xlsr-53 model on the dataset collected by the IDASCO Team
### 0. Setup Environment :
torch
json
codecs
datasets
soundfile
torchaudio
argparse
tokenizers
transformers
speechbrain
pandas
dataclasses
typing
tqdm
numpy
sentencepiece
All the python packages can be installed via `pip install`
### 1. Description
#### 1.2 Dataset
We constructed a corpus from 103 sentences
read by 5 speakers. We randomly selected 11
sentences for the test (2min30s) and 92 remaining sentences for the
training (21min51s)
#### 1.3 running:
- Define the tokenization method:
All: letter + tone , char: tone separate to letter and BPE
`python3 custom_tokenizer.py --dataset train --tokenizer all`
- Running asr
`python3 custom_asr.py --model $model --$tok`