# StepFun releases Step-Audio-EditX, a 3B open-weights model for editing and generating speech

> A 3B model that edits emotion and style in speech and does zero-shot text-to-speech. Weights and code are public; the weights licence is not stated.

Published 2025-11-12

Canonical: https://shipped.anth.us/open-weights/stepfun-step-audio-editx/

StepFun released the weights and inference code for Step-Audio-EditX on 12 November 2025. It is a 3B-parameter model that edits emotion, speaking style and paralinguistic sounds in speech, and it also does zero-shot text-to-speech and voice cloning.

The model card lists Mandarin, English, Sichuanese and Cantonese, with Japanese and Korean added later. It needs an NVIDIA GPU; the card lists 12 GB of memory as the minimum. A 4-bit AWQ build is published for smaller cards. The weights are on Hugging Face and ModelScope, and the code is on GitHub. A technical report is on arXiv.

The repository says its code is under Apache 2.0. The model card does not state a separate licence for the weights, so check before shipping it in a product.

The authors report that it beats two commercial systems at cloning and emotion control. Those are their own evaluations, and we did not rerun them.

It does what the card says. The card says less about the weights licence than about the benchmarks.

## Sources

- [Model card on Hugging Face](https://huggingface.co/stepfun-ai/Step-Audio-EditX)
- [Code on GitHub](https://github.com/stepfun-ai/Step-Audio-EditX)
- [Technical report on arXiv](https://arxiv.org/abs/2511.03601)

## Editor's first pass (AI-assisted)

- Real 4.5: Weights are public and ungated, with inference code and a quantised build.
- Useful 3.5: Editing emotion and style of existing speech is a real task, but it needs a 12 GB GPU.
- Notable 3.5: A credible open option for expressive editing. The comparisons are the authors' own.
- Novel 3.5: Iterative editing of speech attributes is a different framing from plain text-to-speech.