AI / ML Completed
WazobiaVoice
Multilingual TTS & Voice Cloning for Nigerian Languages
Overview
A ~2B-parameter multilingual text-to-speech and zero-shot voice cloning model covering Yoruba, Hausa, Igbo, Nigerian Pidgin, and Nigerian English — one model, five languages.
WazobiaVoice extends Chatterbox's multilingual TTS architecture to support Nigerian languages it wasn't originally designed for — including Yoruba's tonal diacritics. Fine-tuned via LoRA adapters on AMD Developer Cloud (ROCm) across a ~31k-clip, ~14GB training set.
Training ran to completion across 7 epochs, with checkpoints merged and published to HuggingFace via an isolated Kaggle merge workflow (built to work around disk-quota constraints). Along the way this involved fixing a silent double from_pretrained call that was discarding prior epoch progress at merge time, missing --no-deps dependencies, a checkpoint epoch-labeling off-by-one, and a CUDA device-side assert crash resolved via per-epoch subprocess isolation. Evaluated across epochs using DNSMOS scoring.