VoxCPM2 - Open-source multilingual speech synthesis with voice cloning that sounds like real speech

VoxCPM2 is a tokenizer-free TTS model that supports 30 languages including Korean. It can synthesize speech from text alone without language tags, and is designed to target 48kHz studio-grade output.

The voice generation methods are also specific.

• Voice Design: Simply prepending a natural language description like (A young woman, gentle and sweet voice) lets you specify gender, age, tone, emotion, and even speed.
• Controllable Voice Cloning: Clones timbre from a short reference clip while allowing separate control over emotion, speed, and expression.
• Ultimate Cloning: Feeding both reference audio and a script together reproduces timbre, rhythm, emotion, and style with greater precision.

The model has a 2B parameter architecture based on MiniCPM-4, and was trained on over 2 million hours of multilingual speech data. Its structure is described as a LocEnc → TSLM → RALM → LocDiT pipeline.

The runtime burden is also relatively low. According to the HF model card, it can run on about 8GB VRAM, with an RTF of about 0.30 on an RTX 4090, and about 0.13 with Nano-vLLM acceleration.

Fine-tuning is possible with just 5-10 minutes of audio, and both LoRA and full SFT are supported. Derivative implementations are also provided, including CPU inference, ONNX, Apple Neural Engine, a Rust reimplementation, and ComfyUI nodes.

The license is Apache-2.0, so commercial use is also permitted.