Training LID Models Using ECAPA-TDNN from SpeechBrain

pip install -r requirements.txt

Data Preparation

Audio files should be converted into .wav files, and structured as follows:

├── root
    ├── language_x
        ...
    ├── language_y
        ...
    └── language_z
        ...

Audio files should be DIRECTLY under ecah language folder for split_dataset.py to work properly.

Splitting the Dataset

Assuming the dataset is not yet split into train/val/test set, run the following:

python split_dataset.py -d path/to/root/folder -v fraction_of_val_set -t fraction_of_test_set

Create WDS Shards

We follow the recipe of Voxlingua107 from speechbrain and create WDS shards next.

cd lang_id
python create_wds_shards.py -v path/to/train -s path/to/train/destination
python create_wds_shards.py -v path/to/val -s path/to/val/destination 
python create_wds_shards.py -v path/to/test -s path/to/test/destination 

Start Training

Remember to go through lang_id/hparams/train_ecapa.yaml and update the config according to the dataset.

python train.py hparams/train_ecapa.yaml

Testing

The WDS shards are not currently used during testing, and only the metadata and the original files are used. Remember to go through lang_id/test/hyperparams.yaml and update the config as well, especially the label_encoder.

cd test
python test.py -m path/to/test/meta -d path/to/original/test/data