AI audio is moving beyond isolated text-to-speech. The more interesting workflow is scene generation: describing dialogue, ambience, music, pacing, and sound effects as one coordinated production brief.
Seed Audio 1.5 is an upcoming model built around that idea. While it is not generally available in the workspace yet, creators can already explore the scene-first workflow using the live Seed Audio 1.0 generator at SeedAudio15.org.
Availability note: Seed Audio 1.5 is coming soon. The online generator currently runs Seed Audio 1.0. The 1.5 capabilities discussed below are published pre-release specifications and may change at launch.
Why “complete audio” matters
Most audio pipelines are assembled one asset at a time:
- Generate or record dialogue.
- Find background music.
- Add ambience and room tone.
- Place sound effects on a timeline.
- Mix everything so the elements feel like one scene.
That approach works, but it creates friction. Voices may not match the room, music may fight the dialogue, and every revision can mean rebuilding several layers.
A complete-audio model treats the scene as the unit of generation. Instead of asking for a voice clip, you write a compact production brief describing what the listener should hear from beginning to end.
For example:
A quiet late-night radio studio during heavy rain.
Host: warm, close-mic voice, calm pace.
Guest: slightly distant voice, thoughtful delivery.
0:00–0:04 — soft rain against the window and low room tone.
0:04–0:14 — host introduces the topic while a restrained piano motif enters.
0:14–0:24 — guest responds; add one distant thunder roll without masking speech.
0:24–0:30 — dialogue ends, rain and piano fade naturally.
This prompt describes speakers, environment, timing, music, and effects together. Even when the result needs editing, it provides a coherent starting point.
What Seed Audio 1.5 is expected to add
Based on the currently published specifications, Seed Audio 1.5 is expected to support:
- Up to six minutes of generated audio
- Up to six reference audio files
- Text, audio, and video-conditioned generation
- 30 languages and regional variants
- Video-aware dubbing and timestamp control
- Separate tracks for dialogue, music, ambience, and effects
The separate-track capability may be especially important for production. A mixed preview is convenient, but stems give editors room to rebalance dialogue, replace music, or reduce an effect without regenerating the full scene.
Video-conditioned input also changes the workflow. Instead of manually translating visual events into a prompt, a model can potentially use the source video as timing and context for dubbing and sound design.
What you can try today
The workspace at SeedAudio15.org currently uses Seed Audio 1.0. It is useful for learning the prompt structure that complete-audio generation rewards:
- Describe the setting before the dialogue.
- Give each speaker a role, delivery style, and position.
- Put sound events in playback order.
- Keep ambience and music subordinate to speech when intelligibility matters.
- Start with text-only prompts, then add references when you know what should remain consistent.
The live model accepts a text prompt, up to three audio references, or one image reference. Paid Seed Audio 1.0 generations can run up to two minutes, while free previews are intentionally short.
A reusable prompt structure
A practical scene prompt can be organized into five layers.
1. Scene
State the location, time, and acoustic character.
Small train station café at sunrise; reflective windows, light crowd noise,
soft interior reverb.
2. Speakers
Describe roles and delivery without naming real people or protected characters.
Narrator: clear adult voice, measured pace, quietly optimistic.
Barista: brief replies, natural conversational rhythm, slightly off-mic.
3. Timeline
Place dialogue and events in order.
0:00–0:03 — espresso machine hiss and distant train brakes.
0:03–0:12 — narrator begins.
0:12–0:16 — barista replies; cup placed on saucer.
4. Sound layers
Name ambience, music, and effects separately.
Ambience: low café chatter and ventilation.
Music: sparse felt piano, very low under speech.
Effects: one cup placement and one train departure bell.
5. Ending
Specify how the scene resolves.
Let the final piano note ring for one second, then fade with the station ambience.
This structure is detailed enough to guide the model without turning the prompt into an unmanageable screenplay.
Common mistakes
Treating the prompt like a keyword list
“Podcast, rain, emotional, cinematic” leaves the relationships between sounds undefined. Write the prompt in playback order instead.
Overloading a short scene
Too many speakers, effects, and musical changes compete for attention. Start with the essential layers and add complexity in later iterations.
Using references too early
References are helpful for preserving a voice or sonic direction, but they add another variable. Establish a good text-only prompt first.
Confusing the upcoming model with the live model
Seed Audio 1.5 is not live in this workspace yet. Six-minute generation, video input, six references, and separate stems are upcoming capabilities, not controls available in the current Seed Audio 1.0 generator.
Preparing for the release
The best preparation is not waiting for a model dropdown to change. Build a small library of scene briefs now:
- A two-speaker podcast segment
- A product advertisement with music and effects
- A game dialogue scene with environmental ambience
- A multilingual narration test
- A video dubbing brief with explicit timestamps
Testing these structures on Seed Audio 1.0 reveals which instructions are clear, which sound layers compete, and where timing needs more precision. Those lessons should transfer well when Seed Audio 1.5 becomes available.
You can explore the current generator, comparison table, prompt guidance, and early-bird plans at https://seedaudio15.org/.
SeedAudio15.org is an independent resource and is not presented as an official ByteDance website.
Top comments (0)