soniqo/VoxCPM-0.5B-ONNX

VoxCPM 0.5B ONNX

ONNX export of openbmb/VoxCPM-0.5B, the smaller bilingual Chinese/English diffusion-autoregressive TTS model.

This bundle includes the optimized serving graph set from models/voxcpm/export:

Graph	Role
voxcpm-decoder.onnx	unified prefill + token-step graph with shared LM weights stored once
voxcpm-audio-encoder.onnx	reference prompt audio to latent prompt frames
voxcpm-audio-decoder.onnx	generated latents to 16 kHz PCM

The exporter can also emit split voxcpm-text-prefill.onnx and voxcpm-token-step.onnx graphs for isolated debugging, but serving should load voxcpm-decoder.onnx to avoid storing the LM weights twice.

Precision

The autoregressive path is quality-sensitive. This bundle includes FP32 LM graph plus a voxcpm-decoder.fp16w.onnx variant that stores MatMul weights as fp16 while keeping fp32 graph I/O and compute. The audio encoder/decoder stay FP32.

Dynamic activation INT8 is intentionally not used. It can preserve loadability while damaging hidden states and spoken content on this model family.

Validation

Export-time validation on the real openbmb/VoxCPM-0.5B weights:

• PyTorch-vs-ONNX numeric checks passed for the serving graphs.
• fp16-weight storage tensor self-check passed for voxcpm-decoder.fp16w.onnx.

Run a cloned-voice ASR roundtrip before promoting a new published artifact to production.

Source

• Weights: https://huggingface.co/openbmb/VoxCPM-0.5B
• Source: https://github.com/OpenBMB/VoxCPM
• Export scripts: https://github.com/soniqo/speech-models/tree/main/models/voxcpm

Responsible Use

Voice cloning capability is included. Obtain consent before cloning a voice, and do not use the model to impersonate people, mislead listeners, or commit fraud.