Self-contained int4 variant (smallest download: 4-bit llm_decoder weights, dequantized to fp32 at compute → clean on CUDA; audio_embeddings is bf16, lossless). Backbone at repo root + audio_tokenizer/ (Higgs fp32) + voices/. Siblings: -bf16 (default), -fp32.
Bidirectional re-export of k2-fsa/OmniVoice for Sokuji (#351); plain onnxruntime.
License — NON-COMMERCIAL
CC-BY-NC-4.0, inherited from k2-fsa/OmniVoice (Emilia training set). Unofficial re-export, not affiliated with/endorsed by the authors. Some re-uploads mislabel it apache-2.0 — incorrect for the weights; do not use commercially.
