VoxCPM is a tokenizer-free text-to-speech model supporting 30 languages and 48kHz audio output. It enables voice creation from text descriptions and reference-audio voice cloning, runnable locally via CLI, library, or web GUI.
Project overview
The project relies on a tokenizer-free end-to-end diffusion autoregressive architecture to directly output 48kHz studio-quality audio without external upsamplers.
Project type
Audio & Speech
Use cases
Audio & Speech
Deployment
Refer to project documentation
License
Apache-2.0
Best for
Developers, creators, and AI engineers needing multilingual speech synthesis or voice cloning with controllable emotion, pace, and expression.
Key capabilities
Creates a brand-new voice from a natural-language description (gender, age, tone, emotion, pace) without needing reference audio.
Clones voice timbre from a short reference clip and allows style guidance to steer emotion, pace, and expression.
Reproduces vocal nuances by continuing from reference audio and its transcript to preserve timbre, rhythm, emotion, and style.
Outputs 48kHz studio-quality audio directly via AudioVAE V2 with built-in super-resolution.
Supports full SFT and LoRA fine-tuning to adapt to specific speakers, languages, or domains.
Limitations and risks
Voice Design and Controllable Voice Cloning results can vary between runs, impacting reproducibility.
The model officially supports 30 languages; usage with others may require testing or fine-tuning.
There is a potential for misuse in impersonation, fraud, or disinformation.
Getting started
Installation difficulty is rated as easy. It uses a standard pip install command (`pip install voxcpm`) and pre-trained weights are automatically downloaded.
Users can generate audio via CLI using `voxcpm design --text "Hello" --output out.wav` or by using the available Python library APIs. A web GUI is also supported.
Alternatives and comparisons
Provides emotional speech synthesis and zero-shot voice cloning using 30 seconds of reference audio. Supports American and British voices.
A distilled speech recognition model that is 6x faster and 49% smaller for English transcription tasks.
A Python and JavaScript interface for text-to-speech generation using multiple backends like llama.cpp and Hugging Face Transformers.