StepAudio 3 Gen

StepAudio 3 Gen is StepFun’s general audio generation model. It accepts descriptions of a scene and its characters to produce an audio clip containing speech and other sound elements. It belongs to the StepAudio 3 family, but should be evaluated separately from the Realtime, ASR, TTS, and Music products.

Key features

The official demonstration page shows voice design, spoken lines, singing, sound effects, music, and mixed scenes. The technical report describes a unified audio-generation approach, including zero-shot text-to-speech. The useful question for a creator is whether one generated scene can communicate the intended action without requiring every sound to be assembled independently.

For example, you could prototype two characters speaking beside a rainy window, with a quiet background track and a door closing after the last line. Treat this as a suggested test brief, not a sample we generated or a guarantee of exact synchronization.

A practical production workflow

Write a short script and separate it into roles, delivery instructions, and scene events. Describe each speaker’s voice and emotional state consistently. Make the order of events explicit, and identify which background sounds should remain quiet enough for speech to stay intelligible.

Start with one scene. Listen for missing words, speaker changes, unwanted sounds, and timing errors. Revise one instruction at a time so you can tell what improved. Compare several outputs using the same script, and keep the prompt alongside the selected file for later revisions.

Before expanding to a longer project, check how the result fits the picture edit. If you need precise cue points, isolated tracks, or a specific export format, confirm those requirements in the current API documentation rather than assuming a mixed clip provides them. A complete sound scene can still require editing and quality control.

Access and pricing

As checked on September 29, 2026, the official documentation lists stepaudio-3-gen-preview and POST /v1/audio/generate. It labels the preview free for a limited period and says it will be retired when a paid version is introduced. No permanent free tier or future price is promised here. Use the official Voice Studio or documentation to check current access before building an integration.

Limitations and FAQ

Is it an open-weight model? We did not verify downloadable weights or a model license for this release. The existence of a research paper and a public demo does not establish local deployment availability.

Does a Realtime leaderboard result apply to Gen? No. The products serve different tasks. Evaluate the Gen outputs you need rather than borrowing another family member’s score.

Can it replace a sound editor? Test that against your requirements. For final delivery, review pronunciation, levels, timing, and consistency across scenes. We have not independently benchmarked output quality.

Alternatives

Compare voice-focused work with ElevenLabs Turbo v2.5, and music-focused work with MiniMax Music 3. Browse voice AI, models, and StepAudio alternatives.

Next step and sources

Listen to the official samples, read the model documentation, and consult the technical report. Then test one short scene against your delivery checklist. Sources checked September 29, 2026.