TL;DR
VoxCPM2 is a 2B parameter multilingual TTS model that directly generates continuous speech representations without a tokenizer using a diffusion autoregressive architecture, supporting voice design and controllable voice cloning.
Key features
Multilingual synthesis in 30 languages: Direct synthesis of input text without language tags
Voice design: Generate new voices from natural language descriptions (gender, age, tone, emotion, speed, etc.)
Controllable voice cloning: Clone voice from short reference audio clips, adjust emotion/speed/expression with style guides
Ultimate cloning: Faithfully reproduce all vocal nuances (timbre, rhythm, emotion, style) given reference voice and script
48kHz high-quality audio: Output 48kHz studio-quality audio from 16kHz reference input (AudioVAE V2 asymmetric encoding/decoding)
Real-time streaming: RTF ~0.3 on RTX 4090, ~0.13 with Nano-vLLM/vLLM-Omni acceleration
Fully open-source and commercially usable: Apache-2.0 license
When to use it
When building a multilingual speech synthesis system (supports 30 languages)
When you need to design new voices solely from natural language descriptions (voice design)
When you need to clone a specific speaker's voice from short reference audio while controlling style
For TTS applications requiring high-quality (48kHz) audio output