VoxCPM: TTS that synthesizes 30 languages at 48kHz without a tokenizer

VoxCPM2 is a text-to-speech model that directly generates continuous speech representations without discrete tokenization. Based on MiniCPM-4, this 2B parameter model is trained on over 2 million hours of multilingual speech data, resulting in natural intonation and expressiveness. It supports 30 languages and can synthesize speech in the corresponding language using only the input text, without language tags.

1 / 2

The core feature is Voice Design, which creates new voices using only natural language descriptions, without reference audio. Users can specify attributes such as gender, age, emotion, and speed in text to generate brand-exclusive voices. It also supports Controllable Cloning, which replicates a voice from a short reference clip while allowing separate control over emotion and speed. Ultimate Cloning is also possible by providing both reference audio and transcribed text, reproducing even vocal details.

It takes 16kHz input and outputs 48kHz studio-quality audio, handling super-resolution internally without external upsamplers. It supports real-time streaming inference with an RTF of approximately 0.3 on an NVIDIA RTX 4090 and can be deployed as an OpenAI-compatible API via vLLM-based serving. Released under the Apache-2.0 license, it is available for commercial use.

It is suitable for developers who need multilingual content production or customized voices. Unlike existing TTS systems, its architecture does not rely on a tokenizer, enabling more precise voice control. It also supports fine-tuning via LoRA. It provides a Python API, CLI, and web demo for easy integration into various environments.