VoxCPM2 local TTS: voice cloning and design

VoxCPM2 is an open source speech model from OpenBMB. This tutorial covers Python setup, speech generation, voice design from a written description, and cloning with a short recording. The speaker runs the local examples on a Mac M4 Pro and demonstrates 48 kHz WAV output without an external API.

Setup uses a virtual environment with PyTorch, torchaudio, and the VoxCPM package. The speaker calls for a newer Python version, a recent PyTorch build, and a few gigabytes of storage, without specifying exact versions. For a self-hosted service, the presenter describes a vLLM route with an OpenAI-style /v1/audio/speech endpoint, but does not run the GPU host command on the Mac.

The cloning example uses about eight seconds of the speaker's voice to generate a different sentence. Providing the reference transcript can give the model more information about the performance. The speaker reports support for around 30 languages, including Chinese dialects, but says the model does not translate. He judges the French example more successful than the Arabic one.

The comparison considers ElevenLabs, Chatterbox, and Qwen3-TTS. The speaker describes Apache 2.0 licensing and estimates that 8 GB of Nvidia VRAM may suffice to start the model, while a 24 GB card leaves more room for concurrent requests. His recommendation depends on the language, reference audio, prompt, and available hardware.