AI voice model · OpenBMB

VoxCPM2: OpenBMB's Open-Source TTS Model, Explained

VoxCPM2 is a free, Apache-2.0 text-to-speech model that speaks 30 languages, designs voices from a description and clones a voice from a short clip. Self-hosting it needs about 8 GB of GPU memory; if you would rather not run a GPU, SupaVocal is a hosted alternative from $9 for the first month.

  • Model: open source, Apache-2.0
  • 2B parameters, 48 kHz output
  • SupaVocal: no GPU, commercial use
Start creating voiceovers
2B
Parameters in VoxCPM2
30
Languages, plus 9 Chinese dialects
48 kHz
Native audio output
~8 GB
GPU memory to self-host

What is VoxCPM2?

VoxCPM2 is an open-source text-to-speech (TTS) model from OpenBMB, the open-source group behind the MiniCPM language models, released in April 2026. It has 2 billion parameters, is built on a MiniCPM-4 backbone and, as OpenBMB describes it, was trained on more than 2 million hours of multilingual speech. Code and weights are public under the Apache-2.0 license, which allows commercial use.

The model is tokenizer-free. Most modern TTS systems first turn audio into discrete tokens and then predict those tokens like words. VoxCPM2 skips that step: a diffusion autoregressive pipeline (LocEnc, TSLM, RALM, LocDiT) generates continuous speech representations inside the latent space of its AudioVAE V2. The practical result is natural prosody and 48 kHz audio straight out of the model, with no separate upsampler.

One checkpoint covers three jobs: plain text to speech, voice design from a written description, and voice cloning from a reference clip. You can use VoxCPM2 from Python, a command-line tool, a local web demo or an OpenAI-compatible server.

VoxCPM2 specs at a glance

The public VoxCPM2 release as of November 2026.

MakerOpenBMB (institutions listed: ModelBest and THUHCSI)
ReleasedApril 2026
Size2B parameters, bfloat16 weights
ArchitectureTokenizer-free diffusion autoregressive TTS on a MiniCPM-4 backbone
Languages30, including English, Chinese, Japanese, Korean, Spanish, French, German, Arabic and Hindi, plus 9 Chinese dialects such as Cantonese and Sichuanese
AudioTakes a 16 kHz reference clip, outputs 48 kHz
ModesText to speech, voice design, controllable cloning, transcript-guided cloning, streaming
Max sequence8,192 tokens
SpeedReal-time factor about 0.30 on an RTX 4090, about 0.13 with Nano-vLLM
GPU memoryAbout 8 GB
Fine-tuningFull SFT or LoRA, from 5 to 10 minutes of audio
LicenseApache-2.0 for code and weights, free for commercial use

VoxCPM 2 vs VoxCPM1.5 vs VoxCPM-0.5B: which version?

Three generations are public. VoxCPM2 is the one OpenBMB recommends; the older two stay useful on smaller GPUs or for Chinese and English only.

VoxCPM2VoxCPM1.5VoxCPM-0.5B
ReleasedApril 2026December 2025September 2025
StatusLatestStableLegacy
Languages302 (Chinese, English)2 (Chinese, English)
Sample rate48 kHz44.1 kHz16 kHz
Voice designYesNoNo
Controllable cloningYesNoNo
Cloning modeIsolated reference or continuationContinuation onlyContinuation only
GPU memoryAbout 8 GBAbout 6 GBAbout 5 GB
RTF on RTX 4090About 0.30About 0.15About 0.17

A real-time factor (RTF) below 1 means the audio is generated faster than it plays back: at 0.30, one minute of speech takes roughly 18 seconds on that card.

What VoxCPM2 TTS does well

These strengths come from the design, and they set VoxCPM2 TTS apart from older open models.

30 languages, no language tag

Type text in any supported language and VoxCPM2 reads it directly, with no language tag needed.

Voice design from words

Put a description in parentheses before the text, such as a young woman with a gentle voice, and the model invents a matching voice with no reference audio.

Controllable cloning

Clone a timbre from a short clip, then steer pace, emotion or style with an instruction while the voice stays recognizable.

48 kHz audio

Studio sample rate straight from the model, which matters for podcasts, video and music beds.

Fast enough to stream

Streaming generation and an RTF near 0.13 with Nano-vLLM make it usable for live voice agents on one card.

Open and tunable

Apache-2.0 weights, LoRA fine-tuning from minutes of audio, and community ports to ONNX, GGUF, Rust and ComfyUI.

How to run VoxCPM2 locally

The quickest route to VoxCPM2 is the Python package. Plan on a machine with an NVIDIA GPU, or an Apple Silicon Mac for slower runs.

  1. 1

    Check the requirements

    You need Python 3.10 to 3.12, PyTorch 2.5 or newer and CUDA 12.0 or newer, plus about 8 GB of free GPU memory for VoxCPM2 itself.

  2. 2

    Install the package

    Run pip install voxcpm. The first call to VoxCPM.from_pretrained("openbmb/VoxCPM2") downloads the weights; ModelScope is an alternative download source.

  3. 3

    Generate speech

    Call model.generate(text=...) and save the waveform at the model's 48 kHz sample rate. cfg_value 2.0 and 10 inference timesteps are the values used in the official examples.

  4. 4

    Try design and cloning

    Use voxcpm design with a --control description, or voxcpm clone with --reference-audio. For a browser UI, python app.py --port 8808 starts the local web demo.

VoxCPM2 hardware and deployment options

Pick the runtime by where the audio has to be made. Speeds are OpenBMB's published figures.

RouteRuns onBest forSpeed
voxcpm Python packageCUDA GPU, Apple MPS or CPUScripts, notebooks, testingRTF about 0.30 on RTX 4090
Nano-vLLM-VoxCPMNVIDIA GPUMany requests at once, FastAPI serverRTF about 0.13 on RTX 4090
vLLM-OmniNVIDIA GPU, multi-GPU, other acceleratorsProduction, OpenAI-compatible /v1/audio/speechBatched and streaming
llama.cpp-omni (GGUF)CPU, Metal, CUDA, VulkanLaptops and edge devices, no PythonRTF about 1.76 on Apple M4 Pro (Q8_0)

An RTF above 1, as on the M4 Pro, means generation is slower than real time: fine for batch narration, too slow for a live agent.

VoxCPM voice cloning: the three modes

VoxCPM voice cloning comes in three flavors, and choosing the right one decides how close the copy sounds.

  • Controllable cloning takes a reference clip through reference_wav_path. The model copies the timbre and still follows a style instruction, such as slightly faster with a cheerful tone. VoxCPM2 is the first version with this mode.
  • Ultimate cloning adds the clip's exact transcript as prompt_text. The model then continues from the reference, keeping rhythm, emotion and small vocal habits. OpenBMB suggests passing the same clip as both prompt and reference for the highest similarity.
  • Voice design needs no clip at all: the voice comes from the description alone.

Results from design and controllable cloning can vary between runs, so generate one to three takes and keep the best. Fixing the seed makes a good take reproducible.

Whatever the tool, clone only your own voice or a voice whose owner has agreed. OpenBMB forbids impersonation, fraud and disinformation, and recommends labeling AI audio.

VoxCPM2 limitations to plan for

None of these block real projects, but they shape how you use the model.

Do

  • Generate 2 or 3 takes for designed or style-controlled voices
  • Split long scripts into paragraphs; very long or highly expressive inputs can turn unstable
  • Give the transcript of your reference clip when similarity matters most
  • Test each language you ship: quality depends on how much training data it had

Don't

  • Expect identical output twice unless you fix the seed
  • Assume a language outside the 30 will work without fine-tuning
  • Run it on a 4 GB card; budget about 8 GB of GPU memory
  • Clone a voice without the speaker's consent

VoxCPM2 vs other open TTS models

Zero-shot cloning scores on the Seed-TTS-eval benchmark, as published by OpenBMB. Lower WER/CER means fewer pronunciation errors; higher SIM means the clone sounds closer to the reference.

ModelSizeEnglish WER %English SIM %Chinese CER %Chinese SIM %
VoxCPM22B1.8475.30.9779.5
Qwen3-TTS1.7B1.2371.71.2277.0
IndexTTS21.5B2.2370.61.0376.5
CosyVoice20.5B3.0965.91.3875.7
F5-TTS0.3B2.0067.01.5376.0

VoxCPM2 trades a little English accuracy against Qwen3-TTS for clearly higher speaker similarity, which is what you notice most when cloning.

No GPU? A hosted alternative to VoxCPM2

VoxCPM2 is built to be self-hosted. That is the right choice if you need its 30 languages, want to fine-tune on your own data, or must keep audio on your own machines, and the Apache-2.0 license lets you ship the results without a usage fee.

If you just need finished voiceovers and don't want to manage Python, CUDA and a GPU, SupaVocal is a separate, hosted AI voice studio; it does not run VoxCPM2 or any VoxCPM release. You type a script in text to speech, pick one of 30+ English and Chinese voices and download the audio, with commercial use on every plan.

  • Voice cloning: upload a clip of at least 3 seconds in Create Voice, confirm you have the speaker's consent, and save up to 10 voices.
  • Voice design: describe a voice in plain words instead of uploading audio.
  • More tools: speech to text, a voice changer, audio translation and sound effects.

Plans start at $19 a month, $9 for the first month, for 200,000 characters, about 200 minutes of speech. See pricing for every plan. Failed generations refund their credits automatically.

Frequently asked questions

What is VoxCPM2?

VoxCPM2 is OpenBMB's open-source text-to-speech model, released in April 2026. It has 2 billion parameters, speaks 30 languages plus 9 Chinese dialects, outputs 48 kHz audio, and can design a voice from a description or clone one from a short clip. Code and weights are under the Apache-2.0 license.

Is VoxCPM2 free for commercial use?

Yes. Both the code and the weights are released under Apache-2.0, which permits commercial use. You pay only for the hardware you run it on. OpenBMB still forbids impersonation, fraud and disinformation, and recommends testing and a safety review before any production deployment.

How much VRAM does VoxCPM2 need?

OpenBMB lists about 8 GB of GPU memory for VoxCPM2, compared with about 6 GB for VoxCPM1.5 and 5 GB for VoxCPM-0.5B. Without a suitable GPU you can run GGUF builds on a CPU or Mac, but expect slower than real-time output.

Can VoxCPM2 run on a Mac or CPU?

Yes. The Python package supports Apple MPS and CPU through its device setting, and the llama.cpp-omni GGUF build runs on CPU, Metal, CUDA and Vulkan. OpenBMB reports a real-time factor near 1.76 on an Apple M4 Pro with Q8_0 weights, so a minute of speech takes close to two minutes.

How good is VoxCPM voice cloning?

It is among the strongest open options for similarity. On OpenBMB's published Seed-TTS-eval run, VoxCPM2 scores 75.3% English and 79.5% Chinese speaker similarity, ahead of Qwen3-TTS and CosyVoice2. For the closest copy, supply the reference clip's transcript too, which switches it into continuation-based cloning.

What is the difference between VoxCPM 2 and VoxCPM1.5?

VoxCPM2 grows from 2 to 30 languages, raises output from 44.1 kHz to 48 kHz, and adds voice design and controllable cloning with style instructions. VoxCPM1.5 is smaller, needs about 6 GB of GPU memory and runs about twice as fast on an RTX 4090, but supports Chinese and English only.

Does SupaVocal use VoxCPM2?

No. SupaVocal is a separate, hosted voice studio and does not run VoxCPM2 or its older versions. It is an alternative for people who want text to speech, voice cloning and voice design without a GPU: 30+ English and Chinese voices, commercial use on every plan, from $19 a month with the first month at $9.

Is there an online alternative to VoxCPM2 TTS?

Yes. If self-hosting VoxCPM2 TTS is more setup than you need, SupaVocal runs in the browser: paste a script of up to 15,000 characters, choose a voice and download the audio. Cloning needs a clip of at least 3 seconds. Plans start at $19 a month, and failed generations are refunded automatically.

Skip the GPU setup: try SupaVocal instead of VoxCPM2

Type a script, pick from 30+ English and Chinese voices or clone your own from a 3-second clip. $9 for your first month, commercial use included.

Start creating voiceovers