Fine-Tuning MMS Adapter Models for Low-Resource ASR
Scripts and utilities for fine-tuning Meta’s Massive Multilingual Speech (MMS) adapter models for ASR on low-resource languages like Mooré.
Speech Processing · Low-Resource NLP · Hugging Face · ASR · TTS
Python, Speech Processing, TTS, ASR
★ 3 stars on GitHub LinksCode All projects
Overview
Built an open-source toolkit to collect, preprocess, align, and normalize low-resource Mooré language speech/text data. Mooré is spoken by 7M+ people across 5+ West African countries.
Key Features
- Data Collection: Automated scraping of parallel text-audio data from various sources
- Preprocessing: Audio resampling, normalization, and cleaning pipelines
- Forced Alignment: Integration with MMS Forced Aligner for precise text-audio alignment
- HuggingFace Integration: Direct export to HuggingFace Datasets format
Impact
This toolkit helps the research community address data scarcity challenges for low-resource languages and enables downstream NLP applications like:
- Automatic Speech Recognition (ASR)
- Text-to-Speech (TTS)
- Machine Translation