VoxCPM Speech Series

The VoxCPM speech series is the OpenBMB Open-source Community's family of tokenizer-free text-to-speech models (no discrete audio tokenizer)🧑‍🧑‍🧒. The models use diffusion autoregression to directly model continuous speech representations, covering text-to-speech (TTS), zero-shot voice cloning, streaming synthesis and personalized fine-tuning. VoxCPM2 is currently recommended for new projects; VoxCPM1.5 and VoxCPM-0.5B are kept as historical Chinese-English bilingual versions and compatibility entry points~

Series Overview

VoxCPM2 The currently recommended 2B multilingual version. Supports 30 languages, 9 Chinese dialects, 48kHz output, voice design and controllable voice cloning.

VoxCPM1.5 The 800M Chinese-English version. Outputs 44.1kHz, keeps prompt-audio-continuation voice cloning, and supports SFT and LoRA fine-tuning.

VoxCPM-0.5B The original 640M Chinese-English version. Outputs 16kHz; suitable as a historical baseline, for compatibility with older projects, or as a lightweight validation entry point.

Model	Parameters	Languages	Output Sample Rate	Cloning Method	Positioning
VoxCPM2	2B	30 languages, plus 9 Chinese dialects	48kHz	Isolated reference-audio cloning; also supports prompt audio continuation	Default choice, latest version
VoxCPM1.5	800M	Chinese, English	44.1kHz	Prompt audio continuation, requires reference text	Upgrades and compatibility for older Chinese-English projects
VoxCPM-0.5B	640M	Chinese, English	16kHz	Prompt audio continuation, requires reference text	Original baseline and historical compatibility

Recommended Choices

Goal	Recommended Model	Why
Multilingual generation, voice design, controllable voice cloning	VoxCPM2	The only version that supports 30 languages, an isolated reference-audio channel, Voice Design and Style Control.
Chinese-English projects with existing 1.x projects or LoRA assets	VoxCPM1.5	Upgrades to 44.1kHz output and a 6.25Hz LM token rate within the 1.x architecture, offering a fairly smooth migration path.
Reproducing the original behavior or validating the smallest historical checkpoint	VoxCPM-0.5B	Keeps the original 1.0 architecture and 16kHz output; it should not be taken as representative of VoxCPM2's multilingual, controllability or audio quality capabilities.

Capability Boundaries

VoxCPM2's voice design works by adding a natural-language description in parentheses at the start of the target text; its controllable voice cloning can take reference audio alone, optionally combined with a control description. VoxCPM1.5 and VoxCPM-0.5B clone via prompt audio continuation: you must provide both prompt_wav_path and an accurate transcription of that audio, prompt_text. Don't mix up the two types of interface, their required inputs or their expected results.

Please pay attention to voice rights and safety. Before voice cloning, fine-tuning or public release, you must confirm the speaker's authorization and the boundaries of lawful use. Do not use it for impersonation, fraud or spreading disinformation; synthesized audio that is shared publicly should be clearly labeled as AI-generated~

Related Resources

Topic	Link
Code repository and release notes	OpenBMB/VoxCPM
VoxCPM2 model weights	Hugging Face / ModelScope
Official website	voxcpm.com
Model documentation	VoxCPM 2 / VoxCPM 1.5 / VoxCPM-0.5B
Full documentation	VoxCPM ReadTheDocs
Online demo (mainland China)	VoxCPM Demo
Online demo (international)	Hugging Face Spaces
Technical report	VoxCPM2 Technical Report
Demo page and audio samples	VoxCPM2 Demo Page
VoxCPM2 technical community group	Join the official VoxCPM2 technical community group~