OmniVoice Top-5 Multilingual (experimental)
Experimental research checkpoint — not part of the main WaxalNet release suite.
OmniVoice finetuned on a sqrt-hours-resampled pool of the 5 highest-data-volume WAXAL languages, with adaptive per-language CER thresholds, 5000 steps.
Languages: Yoruba (yo), Swahili (sw), Kinyarwanda (rw), Igbo (ig), Hausa (ha)
License note
This model is finetuned from k2-fsa/OmniVoice, whose pretrained weights are licensed CC-BY-NC-4.0 (non-commercial) due to constraints from its Emilia training data. This finetune inherits that restriction: research use only, not for commercial use.
Training recipe
- Base:
k2-fsa/OmniVoice(Qwen3-0.6B backbone + Higgs Audio v2 codec) - Data: WAXAL multilingual corpus, cleaned via the project's Phase 1 pipeline (clip/silence/duration checks + ASR-CER gate), with a stricter CER ceiling than the main VoxCPM2 release models to test whether label-noise reduction improves TTS fidelity.
- For multilingual variants: sqrt-hours resampling across languages, same methodology as the VoxCPM2 Top-5/Top-10 models, with adaptive per-language CER thresholds (floored at 0.08, capped at 0.3, relaxed to each language's own median where the ASR model itself is weak) to avoid collapsing low-resource languages.
- 5000 training steps (bf16, lr 1e-5), OmniVoice's default finetune recipe.
This is a fast research iteration, not a polished release — quality has not been formally evaluated against the project's main VoxCPM2 models.
- Downloads last month
- 27