VoxCPM

VoxCPM2 impresses with its tokenizer-free approach and breadth of capabilities – voice design from text alone, controllable cloning, and support for 30 languages all in one open-source package. For developers who need a customizable TTS engine that runs locally, this is a strong contender. The 2B parameter model produces 48kHz audio that sounds natural and expressive, and the streaming performance (RTF ~0.3 on a 4090) is good enough for real-time applications. Where it falls short is accessibility: there's no hosted API, so you need a GPU with CUDA and some technical chops to get it running. Non-technical users or teams without GPU resources should look at cloud TTS APIs instead. Compared to other open-source TTS like Coqui TTS or Piper, VoxCPM2 offers more languages and a novel tokenizer-free architecture that reduces artifacts. However, the community and documentation are still maturing – you'll need to be comfortable reading GitHub issues and adapting code. In practice, the voice cloning quality depends on the reference audio quality, and the voice design feature can be hit or miss for nuanced descriptions. If you're building a commercial product that needs in-house TTS, the Apache-2.0 license is a big plus. For casual use or rapid prototyping, the setup overhead might be too high.