Moore Speech Corpora Toolkit
Open-source toolkit to collect, preprocess, align, and normalize Mooré language speech/text data for low-resource NLP applications.
Speech Recognition · NLP · Low-Resource Languages · Hugging Face
PyTorch, Hugging Face Transformers, Speech Processing LinksCode All projects
Overview
A comprehensive toolkit for fine-tuning Meta’s Massive Multilingual Speech (MMS) adapter models for Automatic Speech Recognition (ASR), specifically designed for low-resource languages like Mooré.
Features
- Adapter-Based Fine-tuning: Efficient training using adapter layers
- Flexible Data Loading: Support for Hugging Face datasets and local files
- CTC Loss Training: Connectionist Temporal Classification for sequence-to-sequence ASR
- WER Evaluation: Word Error Rate metrics for performance assessment
Repository Structure
1
2
3
4
5
6
7
8
9
10
├── data/ # Data preprocessing scripts
├── models/ # Model loading and checkpoints
├── scripts/
│ ├── preprocess.py # Data preprocessing
│ ├── train.py # Training script
│ ├── evaluate.py # Evaluation with WER
│ └── inference.py # Run inference
├── utils/
│ └── config.yaml # Configuration file
└── logs/ # Training logs
Quick Start
1
2
3
4
5
6
7
8
9
10
11
# Preprocess data
python scripts/preprocess.py --data_source huggingface
# Train model
python scripts/train.py --config utils/config.yaml
# Evaluate
python scripts/evaluate.py --checkpoint models/checkpoints/
# Inference
python scripts/inference.py --audio_file audio.wav
Why Adapter Fine-tuning?
Adapter-based fine-tuning offers several advantages:
- Parameter Efficient: Only train adapter layers, not the full model
- Memory Efficient: Smaller memory footprint during training
- Multi-task Ready: Share base model across multiple languages
- Quick Adaptation: Fast fine-tuning for new languages
About MMS
Meta’s MMS supports 1,100+ languages with models for:
- Automatic Speech Recognition (ASR)
- Text-to-Speech (TTS)
- Language Identification