VoxCPM
Tokenizer-free text-to-speech system generating 48kHz multilingual audio, voice design, and voice cloning.
Generate high-fidelity 48kHz speech across 30 languages with zero-shot voice design and zero-shot cloning.
VoxCPM is an open-source text-to-speech system designed to turn written text into natural multilingual audio. You can generate custom voices from text prompts or clone existing voices from short audio samples. It suits engineering teams building voice safety alerts, automated site briefing systems, or training narration.
What it is
VoxCPM generates continuous speech representations directly via a diffusion autoregressive architecture without discrete tokenization. It produces 48kHz studio-quality audio across 30 languages, supports voice creation from natural language descriptions, and executes voice cloning with prompt guidance.
Who it's for
You should use VoxCPM if you need self-hosted, offline voice generation for construction field alerts, worker safety broadcasts, or multi-language training videos. It is not suitable if you lack GPU infrastructure or need a turnkey non-technical desktop app.
Key features
Tokenizer-Free Audio Generation
Generates continuous speech directly using an end-to-end diffusion autoregressive architecture.
Natural Language Voice Design
Creates custom voices based solely on text descriptions specifying age, tone, emotion, and pace.
High Resolution Audio Output
Outputs native 48kHz studio audio without requiring separate external upsamplers.
Flexible Deployment Engine
Runs on OpenAI-compatible vLLM-Omni servers or C++ edge engines via llama.cpp-omni.
Up and running in 3 steps
Our honest verdict
What's great
- Outputs native 48kHz audio directly from 16kHz reference clips without extra upsampling tools.
- Supports 30 primary languages alongside multiple regional dialects.
- Provides C++ edge deployment options with GGUF quantizations for local hardware.
- Offers full voice creation from simple text prompts without reference audio files.
Where it falls short
- Requires high-end NVIDIA GPU hardware or specialized local C++ compilation to achieve low latency.
- Needs manual model downloading across multi-gigabyte BaseLM and Acoustic weight files.
- Demands technical expertise with CLI, Python APIs, or vLLM deployments rather than offering a GUI.
Pricing & license
VoxCPM is free to download and use under the Apache-2.0 license. Your only expense comes from hosting, hardware execution, and GPU server infrastructure.
Open remote AI-training roles
768 open remote AI-training roles, updated November 14, 2026; most pay $20–$175 an hour. Zekai lists them free; the application goes to the employer’s platform.
Alternatives
Use it, or have us set it up
FAQ
Can I run VoxCPM locally without an internet connection?
Yes, you can load downloaded model weights into local Python scripts or run offline via llama.cpp-omni on CPU or GPU.
Does VoxCPM require transcript files to clone a voice?
Standard controllable cloning requires only a reference audio clip, though providing a transcript enables full detail matching.
