# VoxCPM

> Tokenizer-free text-to-speech for multilingual voice design and cloning

- Category: Multimodal Generation / Audio, Music & Voice
- Language: Python
- Licence: Apache-2.0
- Repository: https://github.com/OpenBMB/VoxCPM
- Website: https://voxcpm.com
- Canonical page: https://ai-tldr.dev/tools/voxcpm/

## In short

VoxCPM is an open-source text-to-speech (TTS) system from OpenBMB. Instead of turning speech into discrete tokens, it directly generates continuous speech representations through an end-to-end diffusion autoregressive architecture, which the team says produces highly natural and expressive audio. It sits in the Audio, Music & Voice category of the Multimodal Generation tools and is open-source, written in Python and licensed Apache-2.0.

## Overview

VoxCPM is an open-source text-to-speech (TTS) system from OpenBMB. Instead of turning speech into discrete tokens, it directly generates continuous speech representations through an end-to-end diffusion autoregressive architecture, which the team says produces highly natural and expressive audio.

The latest release, VoxCPM2, is a 2-billion-parameter model trained on over 2 million hours of multilingual speech and built on a MiniCPM-4 backbone. It supports 30 languages, voice design from a plain-text description, controllable voice cloning, and 48kHz studio-quality output. The weights and code are released under the Apache-2.0 license, so the project is free to use commercially.

## What it does

- Multilingual synthesis across 30 languages plus several Chinese dialects, with no language tag needed in the input text
- Voice Design: create a brand-new voice from a natural-language description (gender, age, tone, emotion, pace) without any reference audio
- Controllable voice cloning from a short reference clip, with optional style guidance to steer emotion, pace, and expression while keeping the original timbre
- Ultimate cloning that reproduces fine vocal detail by continuing from a reference clip paired with its transcript
- 48kHz high-quality output with built-in super-resolution, accepting 16kHz reference audio and needing no external upsampler
- Real-time streaming generation, with much faster inference available through Nano-vLLM or the official vLLM-Omni serving engine

## Getting started

VoxCPM ships as a Python package. Install it with pip, then load a pretrained model and generate audio with a few lines of code, or use the bundled command-line tool. It requires Python 3.10 or newer (below 3.13), PyTorch 2.5.0 or newer, and CUDA 12.0 or newer.

1. Install the package

   Install VoxCPM from PyPI with pip.

```bash
pip install voxcpm
```

2. Generate speech with the Python API

   Load the recommended VoxCPM2 model, generate a waveform from text, and write it to a WAV file.

```python
from voxcpm import VoxCPM
import soundfile as sf

model = VoxCPM.from_pretrained(
  "openbmb/VoxCPM2",
  load_denoiser=False,
)

wav = model.generate(
    text="VoxCPM2 is the current recommended release for realistic multilingual speech synthesis.",
    cfg_value=2.0,
    inference_timesteps=10,
)
sf.write("demo.wav", wav, model.tts_model.sample_rate)
```

3. Design a voice from text or clone one from the CLI

   The voxcpm command-line tool can design a voice from a description or clone a voice from a reference clip.

```bash
# Design a voice from a description, no reference audio needed
voxcpm design \
  --text "VoxCPM2 brings studio-quality multilingual speech synthesis." \
  --output out.wav

# Clone a voice from a reference recording
voxcpm clone \
  --text "This is a voice cloning demo." \
  --reference-audio path/to/voice.wav \
  --output out.wav
```

4. Run the local web demo

   Launch the bundled web app and open it in your browser to try synthesis interactively.

```bash
python app.py --port 8808  # then open http://localhost:8808
```

## When to use it

- Adding natural multilingual voiceovers to videos, audiobooks, and other content across 30 languages
- Designing a custom brand or character voice from a written description when no recording is available
- Cloning a specific voice from a short clip for narration, dubbing, or accessibility tools
- Building real-time voice features into apps using streaming generation and high-throughput serving engines