OpenViSTAI · Agents · Blockchain · Crypto
Submit a project
Menu
Audio & Voice AI

Step Audio Editx

Step-Audio-EditX is an open-source, 3-billion-parameter audio model for iterative speech editing and zero-shot text-to-speech, with repository code for inference and training.

Explore the practical guide ↓

Independent OpenViST project profile, based on linked sources.

Step Audio Editx — OpenViST project profile in Audio & Voice AI.
OpenViST project illustration. See the linked sources for product details.
Understand the project

Practical overview

When it fits

Use it when speech output needs explicit control over expressive qualities, or when a self-hosted model and training code are important to the workflow.

Who it is for

  • Developers building self-hosted speech-generation or audio-editing workflows.
  • Researchers and practitioners exploring controllable emotion, speaking style or paralinguistic speech.
  • Teams interested in adapting or training an audio model using its released SFT, DPO or GRPO code.

Documented uses

  • Generate speech from text using zero-shot voice cloning workflows.
  • Iteratively adjust emotion and speaking style in speech audio.
  • Control paralinguistic cues such as breathing, laughter and sighs.
  • Experiment with pronunciation control, including polyphonic characters.
  • Train or run the model using the provided code and supported vLLM workflows.
Before you choose

Requirements & limits

The code repository is public and specifies the Apache-2.0 license. Local inference requires downloaded model and tokenizer files, Python dependencies, and an NVIDIA GPU with CUDA.

  • The model has 3 billion parameters and lists 12 GB GPU memory as a critical threshold, with 16 GB described as safer.
  • The documented local setup targets Linux and requires Python 3.12 or later, PyTorch 2.9.1 or later, CUDA Toolkit and an NVIDIA GPU.
  • The project offers iterative emotion and style editing, multiple paralinguistic controls and zero-shot TTS; the repository also releases SFT, DPO and GRPO training code.
  • The documentation recommends audio shorter than 30 seconds per inference for optimal performance.
  • The README lists support for Mandarin, English, Sichuanese, Cantonese, Japanese and Korean, with language or dialect tags used in some workflows.

Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.

Source-based practical guide

Set up Step-Audio-EditX for local inference

Prepare the repository, dependencies and model files for local use.

Before you start

  • Linux
  • NVIDIA GPU with CUDA support; 12 GB memory is a critical value and 16 GB is described as safer
  • Python 3.12 or later, PyTorch 2.9.1 or later, and CUDA Toolkit
  • uv and Git LFS

Steps

  1. Clone the Step-Audio-EditX repository and enter its directory.

  2. Install the project environment with uv sync --refresh, then activate the .venv environment.

  3. Install Git LFS and download the Step-Audio-Tokenizer and Step-Audio-EditX model repositories.

  4. Arrange the downloaded tokenizer and model directories under the model download directory.

Expected outcome

The documented repository, Python environment, tokenizer and model files are prepared for the local inference stage.

Guide limitations

  • The setup instructions do not establish that this profile has run or verified the installation.
  • The repository recommends keeping audio under 30 seconds per inference for optimal performance.
Documentation 1 ↗
Evidence and transparency

Sources & review

Official repository ↗
Source record details
Source availability
Active
Sources checked
15 October 2026
Original website title
stepfun-ai/Step-Audio-EditX · Official homepage title · retained in its source language
How we review sourcesSuggest a correction