voxpm2-ncnn

VoxCPM2 NCNN

This repository contains an NCNN export of openbmb/VoxCPM2 for use with the voxcpm2-ncnn C++ runtime.

It is a converted runtime asset package, not a newly trained or fine-tuned model. The model weights keep the same Apache-2.0 license as the upstream VoxCPM2 release.

Model repository: https://huggingface.co/lyrin/voxpm2-ncnn

Runtime source: https://github.com/LiYulin-s/voxcpm2-ncnn.git

Model Details

• Base model: openbmb/VoxCPM2
• Format: fp16 NCNN .param / .bin component graphs
• Task: multilingual text-to-speech
• Output audio: 48 kHz mono PCM, written by the runtime through FFmpeg
• Runtime target: voxcpm2-ncnn
• License: Apache-2.0 for the model assets

VoxCPM2 is a multilingual controllable speech generation model. The upstream release describes support for 30 languages, 9 Chinese dialects, voice design, style-controllable voice cloning, and high-fidelity continuation cloning. This NCNN package targets the modes currently exposed by the voxcpm2-ncnn runtime.

Files

The exported model directory contains the runtime assets and a local LICENSE copy. The repository root keeps an additional LICENSE copy for hosting tools that expect the license at the top level.

• model.json: NCNN component manifest and runtime settings
• *.ncnn.param, *.ncnn.bin: exported NCNN component graphs and weights
• tokenizer.json: tokenizer asset used by the runtime
• LICENSE: Apache-2.0 license text for the model assets

Export-time intermediate files such as TorchScript, PNNX graphs, and generated Python wrappers are intentionally not included.

Usage

Download the model assets into the runtime repository:

huggingface-cli download lyrin/voxpm2-ncnn --local-dir assets/voxcpm2

Smoke-test the NCNN components:

xmake run voxcpm2 -m assets/voxcpm2 --smoke-components
xmake run voxcpm2 -m assets/voxcpm2 --smoke-components --vulkan

Generate speech from text:

xmake run voxcpm2 -m assets/voxcpm2 \
  -t "你好，欢迎使用 VoxCPM2 NCNN。" \
  -o out.wav

Use prompt continuation with prompt audio:

xmake run voxcpm2 -m assets/voxcpm2 \
  -t "这是续写测试。" \
  --prompt "你好。" \
  --prompt-audio prompt.wav \
  -o out.wav

Use reference audio:

xmake run voxcpm2 -m assets/voxcpm2 \
  -t "这是参考音频测试。" \
  --reference-audio reference.wav \
  -o out.flac

The output format is inferred from the -o extension.

Conversion Notes

This package splits VoxCPM2 into NCNN component graphs:

• text embedding
• base and residual decoder steps
• FSQ and projection layers
• DiT estimator
• stop-token head
• AudioVAE encoder and decoder

The runtime uses a page-style KV cache internally while adapting to the current exported decoder-step NCNN graphs.

Limitations

• This is a conversion package; numerical behavior and performance can differ from the upstream PyTorch runtime.
• Not all upstream inference modes are necessarily exposed by the C++ runtime.
• Quality, speaker similarity, latency, and memory use depend on the NCNN build, device, Vulkan driver, and input audio quality.
• Generated speech and voice cloning should be used only with appropriate rights, consent, and safety review.

Attribution

The original VoxCPM2 model is by OpenBMB / ModelBest. Please refer to the upstream VoxCPM2 model card, project repository, and technical report for model architecture, training, evaluation, and intended-use details.

License

The model assets in this repository are released under Apache-2.0, matching the upstream VoxCPM2 release. The license text is included in LICENSE.