MusinguziDenis/Luganda-ASR

## LUGANDA-ASR This repository demonstrates how to fine-tune popular Automatic Speech Recognition (ASR) models for Luganda using the Hugging Face Trainer API. The provided notebooks walk through data loading, preprocessing, and tokenizer creation, followed by model training. The ASR models explored include: * Wav2Vec-XLS-R * Wav2Vec2-Bert * Whisper * MMS Below is a table with details of open-source Luganda speech recognition datasets that can be used for training and evaluation: | Dataset | Number of Hours | Type of Speech | |---------|-----------------| ---------------| |Mozilla Common Voice Dataset Luganda Dataset | 560 | Read | | Fleurs Dataset | 11 | Spontaneous | | The Makerere Radio Speech Corpus: A Luganda Radio Corpus for Automatic Speech Recognition | 155 | Spontaneous | | Yogera | 251 | Read | The code was inpired by several HuggingFace tutorials: * Fine-Tune W2V2-Bert for low-resource ASR with 🤗 Transformers * Fine-tuning XLS-R for Multi-Lingual ASR with 🤗 Transformers * Boosting Wav2Vec2 with n-grams in 🤗 Transformers * Fine-Tune Whisper For Multilingual ASR with 🤗 Transformers Table showing the performance of some of the finetuned ASR models. | Model | Dataset | WER | |-------|---------|-----| | Whisper | CommonVoice version 14 | 29.9 | | XLS-R | CommonVoice version 14 | 27.9 | | W2V-Bert-2.0 | CommonVoice version 14 | 19.3 | ## License This project is licensed under the MIT License - see the LICENSE file for details. ## Citation This was created by Denis Musinguzi to demonstrate how to finetune ASR models. ```@misc{musinguzi2025asr, author = {Denis Musinguzi}, title = {Luganda ASR models}, year = {2025}, publisher = {GitHub}, journal = {GitHub Repository}, howpublished = {\url{github.com } ```