umoqnier/iwstl-low-resource-experiments

Experiments for participation of IWSTL shared task on low resource track: iwslt.org # IWSLT Experiments - Low Resource Track This repository contains the experiments developed for participation in the IWSLT 2026 shared task, specifically in the low resource track. The main objective is to implement automatic speech translation (AST) systems for **Mapudungun (arn)** $\rightarrow$ **Spanish (spa)**. The NVIDIA canary-1B-v2 model is used as a base, applying PEFT techniques through adapters to optimize performance for these languages with limited available data. For CPU prototyping we can use `canary-1b-flash`. ## Architecture ## Folder structure ```mermaid graph TD root[project_root] --> train_py[train.py] root --> utils_dir[utils/] root --> data_dir[data/] utils_dir --> logging_utils_py[logging_utils.py] data_dir --> processors_py[processors.py] data_dir --> dataset_py[dataset.py] data_dir --> datamodule_py[datamodule.py] ``` ### 1. Data Processing - `process_manifets.py`: Cleans JSON manifest files (converts to lowercase, removes punctuation) and allows filtering audio files that exceed a maximum duration (15 seconds by default). - `data_exploration.py`: Notebook/Script for exploring Mapudungun datasets (via HuggingFace). ### 2. Training - `Multi_Task_Adapters.py`: Adapters tutorial provided by NVIDIA. - `canary_experiments.py`: Prototyping notebook that serves as the basis for the main training script. - `canary_script.py`: The main training script. It implements the training logic, data structure (`CanaryMultilingualDataModule`), and model configuration to adapt Canary to the target languages. ### 3. Audio Augmentation - `custom_aumentation.py`: Defines a data augmentation pipeline using `audiomentations` to improve model robustness by adding Gaussian noise, pitch shifting, band-pass filters, and impulse responses (reverberation). ### 4. Evaluation and Inference - `canary_eval.py`: Prototype notebook that served as the basis for the evaluation script. - `canary_eval_script.py`: Model performance evaluation script using standard me …