VoxCPM 2 vs SuperClonica: studio sound or CPU speed
VoxCPM 2 from OpenBMB is a 2B-parameter text-to-speech model released in April 2026. It outputs 48 kHz audio, speaks 30 languages, clones voices and can create a new one from a text description, all under Apache 2.0. SuperClonica runs Supertonic 3 voices; that model is about twenty times smaller, so it is much faster on a CPU but less flexible.
Below is a side-by-side look based on the VoxCPM documentation and our own comparison table.
What VoxCPM 2 can do
VoxCPM is tokenizer-free: it generates speech in the latent space of an audio VAE through a diffusion autoregressive pipeline built on a MiniCPM-4 backbone.
- 30 languages without a language tag, plus nine Chinese dialects such as Cantonese and Sichuanese.
- 48 kHz output from a 16 kHz reference, with built-in super-resolution.
- Voice design from a natural-language description, and controllable cloning.
- Detailed cloning from a reference clip together with its transcript.
- Weights and code under Apache 2.0, described by the authors as free for commercial use.
VoxCPM 2 on CPU and GPU
On an RTX 4090 the authors report a real-time factor of about 0.3, or about 0.13 with Nano-vLLM; the model needs around 8 GB of VRAM. On an Apple M4 Pro through llama.cpp-omni with Q8_0 weights they report about 1.76.
For a desktop 8-core CPU we estimate 2–4× slower than real time, so a 10-second phrase takes 20–40 seconds. The estimate starts from an ONNX export that reports 3.8 with INT8 weights on a Core i3-12300 and assumes a faster CPU gives up to twice the speed. SuperClonica runs 6.4–9.7× faster than real time, about 1.0–1.6 seconds for the same phrase.
Quality and stability
Our table gives VoxCPM 2 and SuperClonica the same quality mark, 4.5 out of 5. For timbre stability VoxCPM 2 gets 3.5 and SuperClonica 5.
The VoxCPM authors note that voice design and controllable cloning results can vary between runs and suggest generating one to three times. For a single trailer that is fine. For a game where one character speaks thousands of lines, a fixed voice file is easier to keep consistent.
When VoxCPM 2 is the better choice
Choose SuperClonica when speech runs on CPUs, offline and in volume, in one of the 31 Supertonic 3 languages. VoxCPM 2 is the better pick in these cases:
- You need Chinese or one of its dialects.
- You want 48 kHz output for ads, music or video.
- You want to design a voice from a description or clone one instantly.
- You have a GPU with around 8 GB of memory or more.