OmniVoice (k2-fsa) finetuned for Swahili at a 5h data budget - companion checkpoint for the Deep Learning Indaba 2026 (RIAD) poster "Tone on a Budget".

Code, the reference-free tone_i2 tone metric, and the recipe: https://github.com/mosesdaudu001/tone-on-a-budget

These weights are a derivative of OmniVoice (k2-fsa, Apache-2.0). At inference the model loads the Higgs-Audio-v2 codec separately; that codec carries the Boson Higgs Audio 2 Community License (non-commercial, output-restriction). This repo contains only the Apache-2.0-derived OmniVoice weights - obtain the codec from its own source.