NNTISpokenLanguageIdentification

Python version

Python 3.11 (latest 3.11.x)

Environment setup

venv + pip

python3.11 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install --extra-index-url https://download.pytorch.org/whl/cu124 -r requirements.txt

or alternatively use the docker image provided by the Tutors.

Credentials

Set these environment variables before running:

export HF_KEY="your_hf_token"
export WANDB_KEY="your_wandb_api_key"

Running training

Reproduce baseline:

python repro_baseline.py

Default run (uses facebook/mms-300m):

python train_model.py

Select a pretrained model:

python train_model.py --model_id facebook/wav2vec2-xls-r-300m

Data augmentation (waveform)

Waveform augmentation is OFF by default. It is only applied when --enable_augmentation is passed.

Enable waveform augmentation:

python train_model.py --enable_augmentation

Control probability of applying augmentation per sample:

python train_model.py --enable_augmentation --augment_prob 0.8

Enable specific augmentations

If you pass any of the specific --enable_* flags, only those enabled transforms will be used (the others are disabled). Example: enable augmentation but only use pitch shift + noise:

python train_model.py --enable_augmentation --enable_pitch_shift --enable_noise

Available flags:

• --enable_gain
• --enable_time_shift
• --enable_speed_perturb
• --enable_pitch_shift
• --enable_noise

Outputs

• Logs are sent to Weights & Biases (W&B).
• The trained model is saved to ./indic-SLID/inprogress.