VoxCPM: Tokenizer-Free TTS for Expressive Multilingual Speech Generation, Voice Design, and True-to-Life Cloning

👋 Contact us on Feishu | WeChat

---

Contents

• Overview
  • Introduction
  • Highlights
  • News
• Performance
• Quick Start
  • Installation
  • Python API
    • Text-to-Speech
    • Voice Design
    • Voice Cloning
    • Style Control
  • CLI Usage
  • Web Demo
• Models & Versions
• Architecture
• Fine-tuning
• Documentation
• Ecosystem & Community
• Risks and Limitations
• Citation

---

Overview

Introduction

VoxCPM is a tokenizer-free Text-to-Speech system that directly generates continuous speech representations via an end-to-end diffusion autoregressive architecture, bypassing discrete tokenization to achieve highly natural and expressive synthesis.

VoxCPM 2 is the latest major release — a 2.3B parameter model trained on 2.36 million hours of multilingual data, now supporting 30 languages, Voice Design, Style Control, and native 48kHz studio-quality audio output. Built on a MiniCPM-4 backbone.

✨ Highlights

• 🌍 30-Language Multilingual — Input text in any of the 30 supported languages and synthesize directly, no language tag needed
• 🎨 Voice Design — Create a voice from scratch via natural-language description, without any reference audio
• 🎭 Style Control — Control speaking style (emotion, pace, etc.) of a cloned voice through text tags
• 🎤 Zero-Shot Voice Cloning — Clone any voice from a short reference clip with high fidelity
• 🔊 48kHz Native Audio — Redesigned AudioVAE V2 with asymmetric encode/decode for studio-quality output
• 🧠 Context-Aware Synthesis — Automatically infers appropriate prosody and expressiveness from text content
• ⚡ Real-Time Streaming — RTF as low as ~0.15 on NVIDIA RTX 4090

🌍 Supported Languages (30)

Arabic, Burmese, Chinese, Danish, Dutch, English, Finnish, French, German, Greek, Hebrew, Hindi, Indonesian, Italian, Japanese, Khmer, Korean, Lao, Malay, Norwegian, Polish, Portuguese, Russian, Spanish, Swahili, Swedish, Tagalog, Thai, Turkish, Vietnamese

News

• [2026.04] 🔥 We release VoxCPM 2 — 2.3B, 30 languages, Voice Design & Style Control, 48kHz audio! Weights | Docs
• [2025.12] 🎉 Open-source VoxCPM1.5 weights with SFT & LoRA fine-tuning. (🏆 #1 GitHub Trending)
• [2025.09] 🔥 Release VoxCPM Technical Report.
• [2025.09] 🎉 Open-source VoxCPM-0.5B weights & Playground. (🏆 #1 HuggingFace Trending)

---

📊 Performance

VoxCPM achieves state-of-the-art results on public zero-shot TTS benchmarks. VoxCPM 2 results are coming soon — below are VoxCPM-0.5B results. See full benchmarks for details.

Seed-TTS-eval

Model	Params	Open	EN WER%⬇	EN SIM%⬆	ZH CER%⬇	ZH SIM%⬆
CosyVoice2	0.5B	✅	3.09	65.9	1.38	75.7
F5-TTS	0.3B	✅	2.00	67.0	1.53	76.0
IndexTTS2	1.5B	✅	2.23	70.6	1.03	76.5
DiTAR	0.6B	❌	1.69	73.5	1.02	75.3
Seed-TTS	—	❌	2.25	76.2	1.12	79.6
VoxCPM-0.5B	0.5B	✅	1.85	72.9	0.93	77.2
VoxCPM 2	2.3B	✅	coming soon

CV3-eval

Model	ZH CER%⬇	EN WER%⬇	Hard-ZH CER%⬇	Hard-EN WER%⬇
CosyVoice2	4.08	6.32	12.58	11.96
IndexTTS2	3.58	4.45	12.8	8.78
CosyVoice3-0.5B	3.89	5.24	14.15	9.04
VoxCPM-0.5B	3.40	4.04	12.9	7.89
VoxCPM 2	coming soon

---

🚀 Quick Start

Installation

pip install voxcpm

Requirements: Python ≥ 3.10, PyTorch ≥ 2.5.0, CUDA ≥ 12.0. See Quick Start Docs for details.

Python API

🗣️ Text-to-Speech

from voxcpm import VoxCPM
import soundfile as sf

model = VoxCPM.from_pretrained("openbmb/VoxCPM2")

wav = model.generate(
    text="VoxCPM 2 brings multilingual support, voice design, and studio-quality audio.",
    cfg_value=2.0,
    inference_timesteps=10,
)
sf.write("output.wav", wav, 48000)

🎨 Voice Design

Create a voice from a natural-language description — no reference audio needed:

wav = model.generate(
    text="(A warm, gentle female voice in her 30s with a calm and soothing tone) "
         "Welcome to VoxCPM 2, the next generation of realistic speech synthesis.",
)
sf.write("voice_design.wav", wav, 48000)

🎤 Voice Cloning

Clone any voice from a short reference clip:

wav = model.generate(
    text="This is a voice cloning demonstration using VoxCPM 2.",
    reference_wav_path="speaker_reference.wav",
)
sf.write("cloned.wav", wav, 48000)

🎭 Style Control

Clone a voice while controlling speaking style with a text tag:

wav = model.generate(
    text="(Speaking slowly with a whispering, mysterious tone) "
         "The secret lies hidden in the ancient library, waiting to be discovered.",
    reference_wav_path="speaker_reference.wav",
)
sf.write("style_control.wav", wav, 48000)

🔄 Streaming API

import numpy as np

chunks = []
for chunk in model.generate_streaming(
    text="Streaming text to speech is easy with VoxCPM!",
):
    chunks.append(chunk)
wav = np.concatenate(chunks)
sf.write("streaming.wav", wav, 48000)

CLI Usage

# Voice design (no reference audio needed)
voxcpm design \
  --text "VoxCPM 2 brings studio-quality multilingual speech synthesis." \
  --output out.wav

# Voice design with style control
voxcpm design \
  --text "VoxCPM 2 brings studio-quality multilingual speech synthesis." \
  --control "Young female voice, warm and gentle, slightly smiling" \
  --output out.wav

# Voice cloning (reference audio)
voxcpm clone \
  --text "This is a voice cloning demo." \
  --reference-audio path/to/voice.wav \
  --output out.wav

# Hi-Fi cloning (prompt audio + transcript)
voxcpm clone \
  --text "This is a voice cloning demo." \
  --prompt-audio path/to/voice.wav \
  --prompt-text "reference transcript" \
  --output out.wav

# Batch processing
voxcpm batch --input examples/input.txt --output-dir outs

# Help
voxcpm --help

Web Demo

python app.py   # then open http://localhost:7860

Full parameter reference, multi-scenario examples, and voice cloning tips → Quick Start Guide | Usage Guide & Best Practices

---

📦 Models & Versions

	VoxCPM 2	VoxCPM 1.5	VoxCPM-0.5B
Status	🟢 Latest	Stable	Legacy
Parameters	2.3B	800M	640M
Audio Sample Rate	48kHz	44.1kHz	16kHz
LM Token Rate	6.25Hz	6.25Hz	12.5Hz
Languages	30	2 (zh, en)	2 (zh, en)
Voice Design	✅	—	—
Style Control	✅	—	—
Reference Cloning	✅ Isolated & Continuation	Continuation only	Continuation only
SFT / LoRA	✅	✅	✅
RTF (RTX 4090)	~0.15	~0.15	0.17
Weights	🤗 HF / MS	🤗 HF / MS	🤗 HF / MS
Technical Report	Coming soon	—	arXiv
Demo Page	Coming soon	—	Audio Samples

See the full Model & Version History for architecture comparison and migration guide.

---

🏗️ Architecture

VoxCPM adopts a tokenizer-free, diffusion autoregressive architecture that models speech in continuous latent space rather than discrete tokens. The system consists of a four-stage pipeline:

1. Local Encoder — Encodes input audio patches into compact local representations
2. Text-Semantic LM — A causal language model (based on MiniCPM-4) that jointly processes text tokens and audio embeddings to capture high-level semantic intent
3. Residual Acoustic LM — Fuses semantic-level and acoustic-level information through concat-projection to model fine-grained acoustic details
4. Local DiT (CFM) — A Conditional Flow Matching diffusion transformer that generates continuous audio latents conditioned on the LM outputs, producing high-fidelity speech patches

The generated latents are decoded by AudioVAE into waveforms. VoxCPM 2 uses AudioVAE V2 with asymmetric encode/decode design (16kHz encode → 48kHz decode) and sample-rate conditioning, enabling native high-fidelity output without post-processing upsampling.

See VoxCPM 2 Model Details for full architecture documentation and version changelog.

---

⚙️ Fine-tuning

VoxCPM supports both full fine-tuning (SFT) and LoRA fine-tuning. With as little as 5–10 minutes of audio, you can adapt to a specific speaker, language, or domain.

LoRA fine-tuning (parameter-efficient, recommended)

python scripts/train_voxcpm_finetune.py
--config_path conf/voxcpm_v2/voxcpm_finetune_lora.yaml

Full fine-tuning

python scripts/train_voxcpm_finetune.py
--config_path conf/voxcpm_v2/voxcpm_finetune_all.yaml

WebUI for training & inference

python lora_ft_webui.py # then open http://localhost:7860

> **Full guide →** [Fine-tuning Guide](https://voxcpm.readthedocs.io/en/dev_2.0/finetuning/finetune.html) (data preparation, configuration, training, LoRA hot-swapping, FAQ)

---

## 📚 Documentation

Full documentation: **[voxcpm.readthedocs.io](https://voxcpm.readthedocs.io/en/dev_2.0/)**

| Topic | Link |
|---|---|
| Quick Start & Installation | [Quick Start](https://voxcpm.readthedocs.io/en/dev_2.0/quickstart.html) |
| Usage Guide & Cookbook | [User Guide](https://voxcpm.readthedocs.io/en/dev_2.0/chefsguide.html) |
| VoxCPM Series | [Models](https://voxcpm.readthedocs.io/en/dev_2.0/models/voxcpm2.html) |
| Fine-tuning (SFT & LoRA) | [Fine-tuning Guide](https://voxcpm.readthedocs.io/en/dev_2.0/finetuning/finetune.html) |
| FAQ & Troubleshooting | [FAQ](https://voxcpm.readthedocs.io/en/dev_2.0/faq.html) |

---

## 🌟 Ecosystem & Community

| Project | Description |
|---|---|
| [**Nano-vLLM**](https://github.com/a710128/nanovllm-voxcpm) | High-throughput GPU serving |
| [**VoxCPM.cpp**](https://github.com/bluryar/VoxCPM.cpp) | GGML/GGUF: CPU, CUDA, Vulkan inference |
| [**VoxCPM-ONNX**](https://github.com/bluryar/VoxCPM-ONNX) | ONNX export for CPU inference |
| [**VoxCPMANE**](https://github.com/0seba/VoxCPMANE) | Apple Neural Engine backend |
| [**voxcpm_rs**](https://github.com/madushan1000/voxcpm_rs) | Rust re-implementation |
| [**ComfyUI-VoxCPM**](https://github.com/wildminder/ComfyUI-VoxCPM) | ComfyUI node-based workflows |
| [**ComfyUI-VoxCPMTTS**](https://github.com/1038lab/ComfyUI-VoxCPMTTS) | ComfyUI TTS extension |
| [**TTS WebUI**](https://github.com/rsxdalv/tts_webui_extension.vox_cpm) | Browser-based TTS extension |

> See the full [Ecosystem](https://voxcpm.readthedocs.io/en/dev_2.0/) in the docs. Community projects are not officially maintained by OpenBMB. Built something cool? [Open an issue or PR](https://github.com/OpenBMB/VoxCPM/issues) to add it!

---

## ⚠️ Risks and Limitations

- **Potential for Misuse:** VoxCPM's voice cloning can generate highly realistic synthetic speech. It is **strictly forbidden** to use VoxCPM for impersonation, fraud, or disinformation. We strongly recommend clearly marking any AI-generated content.
- **Technical Limitations:** The model may occasionally exhibit instability with very long or complex inputs. VoxCPM 2 introduces Voice Design and Style Control, though results may vary.
- **Language Coverage:** VoxCPM 2 supports 30 languages, but performance may vary depending on training data availability for each language.
- **Usage:** This model is released under the Apache-2.0 license. We do not recommend production use without rigorous testing and safety evaluation.

---

## 📚 Citation

If you find VoxCPM helpful, please consider citing our work and starring ⭐ the repository!

```bib
@article{voxcpm2025,
  title   = {VoxCPM: Tokenizer-Free TTS for Context-Aware Speech Generation
             and True-to-Life Voice Cloning},
  author  = {Zhou, Yixuan and Zeng, Guoyang and Liu, Xin and Li, Xiang and
             Yu, Renjie and Wang, Ziyang and Ye, Runchuan and Sun, Weiyue and
             Gui, Jiancheng and Li, Kehan and Wu, Zhiyong and Liu, Zhiyuan},
  journal = {arXiv preprint arXiv:2509.24650},
  year    = {2025},
}

📄 License

VoxCPM model weights and code are open-sourced under the Apache-2.0 license.

🙏 Acknowledgments

• DiTAR for the diffusion autoregressive backbone
• MiniCPM-4 for the language model foundation
• CosyVoice for the Flow Matching-based LocDiT implementation
• DAC for the Audio VAE backbone

Institutions

ModelBest     THUHCSI

⭐ Star History