VoxCPM

Overview

VoxCPM is a tokenizer-free text-to-speech model from OpenBMB. An end-to-end diffusion autoregressive design on the MiniCPM-4 backbone generates continuous speech representations directly from text, and hierarchical language modeling keeps semantic and acoustic information apart. This page runs VoxCPM 1.5, which OpenBMB lists at about 0.8B parameters, with a 44.1 kHz audio VAE and a training corpus of 1.8 million hours of Chinese and English speech.

OpenBMB calls this context-aware speech generation. The model reads the text to decide how it should sound and adjusts the speaking style to the content, even when no reference voice is supplied.

Choosing a variant: VoxCPM2 is the newer 2B release with 30 languages, 48 kHz output, voices designed from a written description and cloning without a transcript, served as a streaming session. VoxCPM fits English or Chinese work where you already hold a reference recording and its transcript, and you want one finished file per request.

What you can build

• Support lines, IVR menus and notification audio in English or Chinese.
• A narrator voice cloned from a few seconds of an existing recording, reused across new scripts.
• English and Chinese versions of the same explainer video or lesson.
• Story and script readings with no reference clip, where the delivery comes from the text itself.

Inputs

• Prompt (required): the text to read. Text normalization is on, so numbers and abbreviations are written out as words before synthesis.
• Reference Audio Text and its audio clip (optional, as a pair): upload a short clip of the target voice in the audio field and type an exact transcript of it. Provide both or leave both empty. VoxCPM continues from the clip and picks up its timbre, accent, emotional tone, rhythm and pacing.
• Cfg Value (default 2.0, up to 20): higher values follow the prompt more closely, though pushing it too far can reduce quality.
• Inference Steps (default 10, up to 50): more steps generally give a better result.

Outputs

One mono MP3 at 44.1 kHz per request. The two samples on this page are each about 12 seconds long.

Recommended settings

• Start from the defaults. If the voice sounds strained or odd, lower Cfg Value. For short sentences that need clearer delivery, raise it slightly; long passages tend to stay more stable at a lower value.
• Use a clean reference clip. Background noise and room sound in the clip are reproduced along with the voice, and output quality follows the quality of the reference.
• Split long scripts into paragraphs and generate them one by one.

Limitations

• OpenBMB notes that output can become unstable on very long or highly expressive input.
• There is little direct control over emotion or speaking style. Both come mostly from the text and the reference clip.
• Training data is mostly Chinese and English, so other languages are not supported reliably.
• Results can contain artifacts or come out differently from what you expected.

Safety & compliance

The weights and code are released under Apache-2.0. OpenBMB forbids any illegal or unethical use, names deepfakes, impersonation, fraud and disinformation as risks of voice cloning, and recommends marking shared output as AI-generated. The model is described as a research and development release, and OpenBMB advises rigorous testing and safety evaluation before production or commercial use. Clone a voice only with its owner's consent.