cobrayyxx/Bemba-IWSLT2025

Code for IWSLT 2025 purpose # Collection of Code for IWSLT 2025, Low Resource Track, Bemba-English This repo is a collection of a bunch of codes/notebooks that are purposed to create a speech-translation systems, Bemba to English. The directory structure of this repo is like below. ``` . ├── data-related_notebooks ├── evaluation_notebooks ├── files ├── other_notebooks ├── training_notebooks ├── readme.md └── .gitignore ``` ## `data-related_notebooks` directory This repository is consists of some notebooks that are related to preprocessing and exploration of the data, like doing EDA for dataset, augmenting dataset, etc. ## `evaluation_notebooks` directory This repository contains several notebooks that are related to evaluating the model/systems. The models/systems that we evaluated are machine-translation(MT), The whole Cascaded Systems, and The whole End-to-End Systems. BLEU, ChrF++, and AfriCOMET(COMET) are the metric evaluation that we used. For the test-set, we used `devtest` split from FLORES-200 and `test` split from Big-C. ## `files` directory In files directory contains several files, most of them are the augmented dataset files, the transcription and translation files. ## `other_notebooks` directory This directory have two notebook files inside of it. One notebook is for predicting the test-set from the IWSLT organizer and the other is for find the best hyperparameter values using optuna. ## `training_notebooks` directory All of the training scripts and notebooks are placed in this directory. Starting from training machine-translation models, transcription/automatic speech recognition (ASR) model, to training End-to-End model. There are several notebook files in here, some of them were used for training the same base-model with different dataset and some of them were trained on the different base-model. ## Citation If you use this repository or any part of it in your work, please cite the following paper: ``` @inproceedings{hazim-al-farouq-etal-2025-bemba, title = "{B}emba …