Speech Recognition for Tuvaluan
The first automatic speech recognition system for Tuvaluan — a Polynesian language with no prior ASR support. Built by fine-tuning Facebook's MMS model using Samoan as a linguistic bridge.
Record speech in Tuvaluan using your microphone, or upload an audio file. The model transcribes in real time.
Tuvaluan is an Austronesian language spoken by approximately 11,000 people in Tuvalu, a small island nation in the central Pacific. It had no existing speech recognition technology before this project.
Samoan and Tuvaluan share significant phonological and lexical overlap. We fine-tune Facebook's MMS model (pre-trained on 1,100+ languages including Samoan) using CTC-aligned Tuvaluan parliament recordings.
Training data comes from the 2024 Tuvaluan Parliament sessions (June & December). Audio is aligned to official Hansard transcripts using a CTC forced alignment pipeline.
The full alignment, training, and evaluation pipeline is open source. Model weights and training code are available on GitHub.
Built on facebook/mms-1b-all with the Samoan adapter as initialization.
CTC fine-tuning with SpecAugment and speed perturbation, trained on RunPod A40 GPUs.
Beam search decoding uses a word-level 5-gram KenLM language model built from the
full parliament transcripts (~290k words).