VoxCPM2 — Tokenizer-Free Multilingual TTS#

A tokenizer-free TTS from OpenBMB based on a diffusion autoregressive architecture. 2B parameters, trained on 2M+ hours, 30 languages, 48kHz output, built on MiniCPM-4.

Architecture: LocEnc → TSLM → RALM → LocDiT, AudioVAE V2 asymmetric 16kHz→48kHz.

Installation#

bash

pip install voxcpm  # Python ≥3.10, PyTorch ≥2.5, CUDA ≥12
model = VoxCPM.from_pretrained("openbmb/VoxCPM2", device="auto")  # cuda→mps→cpu
# torch.compile issues: optimize=False; HF mirror: export HF_ENDPOINT=https://hf-mirror.com

Models#

V2 (2B)1.5 (0.8B)0.5BSample Rate48kHz44.1kHz16kHzLanguages302(zh/en)2(zh/en)Voice Design✅——VRAM/RTF~8GB/0.306GB/0.155GB/~0.17

30 languages: Chinese, English, Japanese, Korean, French, German, Spanish, Italian, Russian, Arabic, Hindi, Thai, Vietnamese, Turkish, Dutch, Finnish, Norwegian, Swedish, Danish, Polish, Portuguese, Greek, Hebrew, Indonesian, Malay, Burmese, Khmer, Lao, Swahili, Tagalog + 9 Chinese dialects (Sichuan, Cantonese, Wu, Northeastern, Henan, Shaanxi, Shandong, Tianjin, Minnan)

Python API#

python

from voxcpm import VoxCPM; import soundfile as sf
model = VoxCPM.from_pretrained("openbmb/VoxCPM2", load_denoiser=False)

# Basic TTS
sf.write("out.wav", model.generate("Hello!", cfg_value=2.0, inference_timesteps=10), model.tts_model.sample_rate)

# Voice Design (text description → voice, no reference audio needed)
wav = model.generate("(A young woman, gentle voice)Hello!")

# Controllable Cloning (reference audio + style control)
wav = model.generate("Hello.", reference_wav_path="voice.wav")
wav = model.generate("(faster, cheerful)Hi.", reference_wav_path="voice.wav")

# Ultimate Cloning (reference audio + transcript for full detail reproduction)
wav = model.generate("Text.", prompt_wav_path="ref.wav", prompt_text="transcript", reference_wav_path="ref.wav")

# Streaming
import numpy as np
wav = np.concatenate([c for c in model.generate_streaming("Streaming!")])

generate() params: text(required) reference_wav_path prompt_wav_path prompt_text cfg_value=2.0(1-3) inference_timesteps=10(4-30) normalize=False denoise=False retry_badcase=True

CLI#

bash

voxcpm design --text "Hello" --control "Young female warm voice" --output out.wav --device auto
voxcpm clone --text "Hi" --reference-audio voice.wav --prompt-audio ref.wav --prompt-text "txt" --output out.wav
voxcpm batch --input examples/input.txt --output-dir outs

Web Demo#

bash

git clone https://github.com/OpenBMB/VoxCPM.git && cd VoxCPM && pip install -e .
python app.py --port 8808 --device auto

Deployment#

vLLM-Omni (recommended, OpenAI-compatible)#

bash

uv pip install vllm==0.19.0 --torch-backend=auto
git clone https://github.com/vllm-project/vllm-omni.git && cd vllm-omni && uv pip install -e .
vllm serve openbmb/VoxCPM2 --omni --port 8000
curl http://localhost:8000/v1/audio/speech -H "Content-Type:application/json" -d '{"model":"openbmb/VoxCPM2","input":"Hello!","voice":"default"}' --output out.wav

Nano-vLLM: pip install nano-vllm-voxcpm (RTF ~0.13 vs standard ~0.30)

Fine-tuning#

bash

# LoRA (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
python lora_ft_webui.py  # http://localhost:7860

Data format JSONL: {"audio":"path","text":"transcript","ref_audio":"path"} (recommend 30-50% samples with ref_audio). LoRA params r=32 alpha=16, hot-swappable ( load_lora/ unload_lora/ set_lora_enabled). Adapt to a speaker with as little as 5-10 minutes of audio.

License#

Apache 2.0 — free for commercial use