openbmb / VoxCPM2
VoxCPM2
A real-time, multilingual text-to-speech system offering expressive voice design and high-fidelity voice cloning through low-latency streaming inference.
# VoxCPM2
A real-time, multilingual text-to-speech system offering expressive voice design and high-fidelity voice cloning through low-latency streaming inference.
## Overview
VoxCPM2 is a 2B-parameter, diffusion-autoregressive speech model that produces natural, 48 kHz audio across 30+ languages. It supports three complementary modes out of the box:
- Voice Design — describe a voice in natural language (e.g. *"warm female voice, mid-thirties, calm"*) and the model synthesises a matching speaker.
- Voice Cloning — provide a short reference recording and the model reproduces that speaker's timbre on arbitrary text.
- Style Continuation — pair a reference recording with its transcript to carry over the speaker's prosody, pacing, and emotion into the output.
## Use Cases
- Real-time voice agents and conversational assistants
- Long-form narration, audiobooks, and podcasts
- Dubbing, localisation, and cross-lingual voiceover
- Character voices for games and interactive media
- Accessibility tools and screen readers
## Highlights
- Faster-than-real-time generation (RTF ≈ 0.30, or ≈ 0.13 with acceleration)
- Clean, broadcast-quality 48 kHz output
- Robust multilingual coverage, including code-switched text
- Built-in denoising for reliable cloning from noisy inputs
- Streaming API that yields audio as it is generated
Example prompts
Great starting points for VoxCPM2.
API quick start
Run VoxCPM2 with a single API call.
{
"prompt": "A small robot learned to paint by watchin…",
"promptAudioText": "...",
"cfgValue": 2.0,
"inputAudio": "https://your-cdn.com/input.mp3"
}