nexusclouduk/omnivoice-onnx-kv-b1-fp16

OmniVoice WebGPU FP16 KV B1 ONNX Bundle

This repository contains the model bundle used by OmniVoice Vulkan Bridge. The main diffusion model is the fp16, KV-cached, B=1 split ONNX export.

Bridge code: MarkShark2/omnivoice-vulkan-bridge

Model Notes

• —Base model: k2-fsa/OmniVoice
• —Main graph: fp16, KV-cached, B=1 split conditional/unconditional CFG passes
• —Decoder: WebGPU-compatible graph surgery for ONNX Runtime Web 1.25 prerelease
• —Encoder: WebGPU-capable voice-reference encoder with WASM fallback in the bridge

Attribution

Based on OmniVoice by k2-fsa and Xiaomi Corp.

The original inference code for this project was based on Magkino/vocoloco_tts.