- Developers
- ModelBest and THUHCSI
- Architecture
- Continuous speech / diffusion autoregressive
- VoxCPM 1.5
- Chinese/English, 44.1 kHz
- VoxCPM2
- 30 languages, 48 kHz
Strengths worth exploring
Let text context shape delivery
The research separates semantic and acoustic modeling while preserving continuous speech detail. Context can help shape prosody instead of giving every passage the same reading pattern.
Official source [2]Version 1.5 improves bilingual audio detail
Version 1.5 raises output to 44.1 kHz, reduces the internal language-model token rate, and supports SFT and LoRA customization for existing Chinese/English workflows.
Official source [3]Version 2 adds design and controllable cloning
VoxCPM2 expands to 30 languages and 48 kHz output, with natural-language voice design and controllable cloning. These should not be attributed to 1.5.
Official source [5]Different references for different goals
Version 2 can clone from a reference clip alone, or use a reference plus an accurate transcript for audio-continuation-style cloning that carries more of the original rhythm and style.
Voice design creates a voice from text alone. Controllable cloning adds performance direction while retaining a reference speaker.
Official source [1]Pin the checkpoint and use its sample rate
VoxCPM-0.5B, VoxCPM1.5, and VoxCPM2 differ in output rates, languages, and controls. Set the checkpoint explicitly and write audio with model.tts_model.sample_rate rather than a rate copied from an older example.
Official source [4]Usage and deployment
- Prepare supported Python, PyTorch, and compute. The official quick start offers pip installation and documents device-specific requirements.
- Install voxcpm and explicitly load openbmb/VoxCPM2. The minimal example below synthesizes text; choose openbmb/VoxCPM1.5 and its matching API for a 1.5 workflow.
- Validate local short-text generation before starting the official WebUI or following the documented streaming, batch, and serving options.
python -m venv .venv
source .venv/bin/activate
pip install voxcpmimport soundfile as sf
from voxcpm import VoxCPM
model = VoxCPM.from_pretrained("openbmb/VoxCPM2", load_denoiser=False)
wav = model.generate(text="Hello, welcome to the voice model library.")
sf.write("output.wav", wav, model.tts_model.sample_rate)Before you get started
- 44.1 kHz and 48 kHz describe sample rates, not guaranteed fidelity with poor references or unsuitable content.
- Version 2's multilingual, voice-design, and controllable-cloning features should not be listed as version 1.5 capabilities.
Where it fits
Useful for developers exploring detailed speech, contextual delivery, and self-hosting while choosing clearly between bilingual and newer multilingual releases.
Official resources and sources
- OpenBMB/VoxCPM · READMEhttps://github.com/OpenBMB/VoxCPM
- VoxCPM · Technical paperhttps://arxiv.org/abs/2509.24650
- VoxCPM 1.5 · Version guidehttps://voxcpm.readthedocs.io/en/latest/models/voxcpm1.5.html
- VoxCPM · Quick starthttps://voxcpm.readthedocs.io/en/latest/quickstart.html
- VoxCPM2 · Technical reporthttps://arxiv.org/abs/2606.06928
DS Speech currently provides Qwen-Audio-3.0-TTS online. Other models in this library are documented for learning and self-deployment.