Skip to content
Step-Audio-EditX
Zero-shot voice cloning and expressive editing of emotion, style, and paralinguistic cues
Features
System Requirements
32GB RAM recommended. 20GB+ storage recommended.
Windows 10/11: NVIDIA GPU 12GB+ VRAM required.
Note: For NVIDIA GPUs, install a newer driver.Introduction
Step-Audio-EditX is an open-source audio generation and editing large model developed by the StepFun AI team. Designed for expressive and iterative speech editing, it empowers even non-technical users to easily modify voice characteristics—simply provide a short reference audio clip and a text instruction, and the model can precisely adjust emotion, speaking style, or paralinguistic cues.
The complete list of officially supported tags: "Step-Audio-EditX Usage Tips and Installation Guide"
🎯 User-Friendly Features
- Zero-Shot Voice Cloning (TTS): Upload any short audio sample as a “voice template,” and synthesize new text in that exact voice. Supports Mandarin, English, Sichuanese, Cantonese, and more.
- Emotion Editing: Add tags like
[Happy],[Sad], or[Angry]to instantly infuse the desired emotion into the synthesized speech. - Speaking Style Control: Choose from dozens of styles—including whisper, childlike, serious, exaggerated, or act_coy—to make speech more vivid and natural.
- Paralinguistic Editing: Precisely insert human-like nuances such as breathing, laughter, surprised “oh!”, hesitant “uhm”, or sighs for lifelike expressiveness.
⚙️ Technical Architecture & Advantages
Built on a 3-billion-parameter Audio Large Language Model (Audio LLM), Step-Audio-EditX integrates three core components:
- Dual-Codebook Audio Tokenizer: Efficiently converts raw audio into discrete symbolic sequences.
- Audio LLM: Generates new token sequences conditioned on both text instructions and reference audio.
- Flow Matching-Based Audio Decoder: Reconstructs high-fidelity, natural-sounding waveforms from token sequences.
LicenseApache-2.0