Guide

How to prompt Seed-Audio 1.0.

Seed-Audio reads direction, not keywords. The single biggest quality jump comes from writing your prompt like a cue sheet: separate lines for who speaks, how the room sounds, and when the music moves. This guide is the grammar we use for every sample on this site.

1. Structure the prompt as a cue sheet

You can write free prose and the model will cope. It performs noticeably better when the prompt is organized into labelled intents:

[VOICE]A weary detective, low and dry. A younger officer, quick and nervous.

[PACE]Slow, deliberate. Leave a long pause after every question.

[AMBIENCE]Interrogation room: fluorescent hum, distant traffic, a chair creak. Keep it under the voices.

[MUSIC]Sparse double bass. Enters on the second question, gone before the last line.

[SCRIPT]“Where were you at midnight?”
“I already told you. Twice.”

// the exact grammar the model reads. No timeline editing afterwards

[VOICE] describes each speaker: age, texture, mood. Two or three adjectives beat a paragraph.
[PACE] sets tempo and pauses. The model respects explicit pause direction well.
[AMBIENCE] is the room: what hums, what passes by, how far away. Tell it to stay under the voices.
[MUSIC] works best with entry and exit cues, like a real score note.
[SCRIPT] is the exact dialogue, in quotes, one line per speaker turn.

2. Respect the real limits

Each take renders up to 2 minutes of audio. Plan scenes, not chapters.
Prompts accept up to 2,048 characters, but keep voiced text under about 400 characters per take. Beyond that the model speeds up delivery to fit, and it shows.
Languages: English and Chinese. Regional accents respond well to plain description.
No streaming. Every take is a full render, so iterate in short takes.

3. Three prompts you can paste right now

These are the exact prompts behind the samples on the homepage:

Two street vendors argue playfully over the last crate of
mangoes at a busy night market. Lively crowd chatter and
sizzling food carts behind them. The older vendor laughs
first and gives in.

A 1940s radio drama cold open. A low, gravelly male narrator
speaks over faint vinyl crackle and a slow, smoky jazz
trumpet: "The city sleeps. Somewhere out there, a phone is
ringing that should not be."

Steady rain falling on a tin roof, distant soft thunder, and
a lone melancholy cello playing a slow melody. No speech,
just the scene.

4. Common failure modes and fixes

Music drowns the dialogue: add “keep the music under the voices” and give it an exit cue.
Voices drift between lines: describe each voice once, clearly, and keep the cast to 2 or 3 per take.
Rushed delivery: your voiced text is too long for the take. Cut the script or split the scene.
Flat emotion: direct the feeling per line inside the script, like a stage note in brackets.

5. Go render something

The playground runs the real model with free takes. Start from a template, change one cue at a time, and you will internalize the grammar in ten minutes. For model internals and access routes, see the Seed-Audio 1.0 guide.