VoxCPM

Name: VoxCPM
Author: OpenBMB (Tsinghua University)

Audio & Speech

OpenBMB (Tsinghua University)

Open SourceFree

About

OpenBMB's tokenizer-free Text-to-Speech system: directly generates continuous speech representations via diffusion autoregressive architecture. 2B parameter model, 30 languages, Voice Design, controllable voice cloning, 48kHz studio-quality audio.

Visit

Share this tool

Twitter LinkedIn Facebook

Our Verdict

Worth Trying

A tokenizer-free TTS breakthrough from OpenBMB — 30 languages, voice design, and 48kHz output, but GPU-heavy.

VoxCPM2 represents a significant advance in open-source TTS: its tokenizer-free diffusion autoregressive architecture produces more natural speech than discrete-token alternatives, and the single model covering 30 languages with automatic detection is impressive. The Voice Design feature (create a voice from a text description) and Ultimate Cloning (reproduce every nuance from reference audio + transcript) are genuinely useful capabilities.\n\nThe caveats: real-time inference needs an RTX 4090-class GPU, documentation is scattered, and the 101 open issues suggest ongoing development. For researchers, content creators, and developers who need high-quality multilingual TTS with voice design capabilities, VoxCPM2 is a compelling open-source option.

Best for

•Researchers, content creators, and developers needing high-quality multilingual TTS with voice design and cloning, who have access to a powerful GPU (RTX 4090 or better).

Consider alternatives if

•If you lack a powerful GPU, consider cloud-based TTS APIs (ElevenLabs, OpenAI TTS) or lighter open-source models that run on CPU.

Try VoxCPM

Supported Platforms

Web AppWindowsmacOSLinuxAPI

Available platforms include Web App, Windows, macOS, Linux, and API.

Key Features

Tokenizer-free architecture: directly generates continuous speech representations via diffusion autoregressive model

2 billion parameter model (VoxCPM2) trained on 2M+ hours of multilingual speech data

30 languages supported — no language tag needed, auto-detects from input text

Voice Design: create a brand-new voice from natural-language description (gender, age, tone, emotion, pace)

Controllable Voice Cloning: clone any voice from a short reference clip with style guidance

Ultimate Cloning: reproduce every vocal nuance from reference audio + transcript

48kHz studio-quality audio output via AudioVAE V2 with built-in super-resolution

Real-time streaming: RTF as low as ~0.3 on RTX 4090, ~0.13 with Nano-vLLM acceleration

Context-aware synthesis: automatically infers prosody and expressiveness from text content

Built on MiniCPM-4 backbone from OpenBMB

Pricing

free

Free and open source (Apache 2.0). Model weights available on Hugging Face and ModelScope.

Use Cases

Multilingual TTS for content creation, dubbing, and accessibility

Voice design for virtual assistants, gaming characters, and brand voices

High-fidelity voice cloning for personalized speech applications

Pros

Tokenizer-free architecture produces more natural, expressive speech than discrete-token TTS

30 languages in one model with automatic language detection

Voice Design from natural language description — no reference audio needed

48kHz studio-quality output with built-in super-resolution

Apache 2.0 licensed, from the reputable OpenBMB team (Tsinghua University)

Cons

Requires significant GPU resources (RTX 4090 recommended) for real-time inference

101 open issues at verification time

Documentation is scattered across GitHub, ReadTheDocs, and Hugging Face

Last push was 35 days ago — moderate activity level

Latest Update

July 2026: VoxCPM2 released with 2B parameters, 30 languages, Voice Design, controllable cloning, 48kHz output, and real-time streaming.

Related Audio & Speech Tools

ElevenLabs

Leading AI voice synthesis and cloning platform with multi-language support

Whisper

OpenAI

OpenAI's open-source speech recognition model for multi-language speech-to-text

Fish Audio

Voice generation platform from the team behind the open-source TTS star Fish Speech. Clone a voice from just 10-30 seconds of audio; the S1/S2 models deliver natural, expressive speech with commercial use and pay-as-you-go API

MiniMax Audio

MiniMax

MiniMax's voice generation platform. The Speech model family delivers hyper-realistic TTS in 40+ languages with 10-second voice cloning, controllable emotion and sound-effect tags, and a free web trial