Audio Prompting Guide — Write Prompts Like a Screenwriter
Seed Audio 1.0 turns one prompt into dialogue, music, ambience and effects. The prompts that work read like tiny screenplays. This guide shows the structure, the reference syntax, and eight copy-ready examples.
Open the workbenchThe script-style structure: who, feeling, scene, sound
Every reliable scene prompt answers four questions. WHO is speaking — name each character and give them a voice ("male, gravelly, tired"). WHAT they feel — emotion drives the performance more than any adjective about audio quality. WHERE it happens — the space defines reverb, distance and background ("a stone cavern" vs "a car interior"). WHAT ELSE sounds — call out music, ambience and effects explicitly, or the model has to guess.
Write dialogue lines in quotes, in the language you want spoken. Keep each generation to one coherent beat of a story — two to four characters, one location — and let Extend carry the plot forward.
@Audio1–@Audio3: voice cloning by reference
Upload up to three reference clips alongside your prompt, then assign them by writing @Audio1, @Audio2 or @Audio3 wherever a voice is cast: "The detective speaks with @Audio1's voice." Each reference clones the timbre of one voice, so three clips can cast three distinct characters in a single scene.
Clean, dry recordings clone best: a quiet room, no music, a few natural sentences. The clone follows your prompt's emotional direction — a calm reference can still deliver an angry line.
Character images — the other way to cast (mutually exclusive)
Instead of reference audio, you can attach one character image. The model reads age, build, attitude and setting from the picture and invents a fitting voice — great when you have concept art but no voice actor. The two casting methods are mutually exclusive: a prompt uses @Audio references or one image, never both.
Directing TTS with voice_instruction
Text to speech (Seed Speech 2.0) separates the script from the direction. The script is read verbatim; voice_instruction is a private note to the voice — tone, emotion, pacing, attitude — and is never spoken. "Read like a flight attendant announcement", "barely holding back laughter", "grave documentary gravitas": same script, radically different takes.
Combine it with the mechanical controls — speed, pitch, volume — for the final polish. Directing lives in the instruction; physics lives in the sliders.
Extend: the strategy for anything longer than two minutes
Scene generations run up to two minutes. Extend continues an existing piece with the same character voices, music bed and mix — so plan long work as a chain: write beat one, generate, audition, then Extend with a prompt that only describes what happens next. Do not re-describe the cast or the room; the continuation inherits them.
Iterate cheaply at the head of the chain: fixing the opening beat before extending saves regenerating everything downstream. For audiobooks, extend chapter by chapter; for ambience, extend into long seamless loops.
Eight copy-ready example prompts
Common mistakes checklist
- Vague sound words. "A door sound" could be forty different doors — name the material, the space, and the energy: "heavy iron gate creaking open in a stone corridor".
- Putting stage direction inside a TTS script. In text to speech, everything in the script gets read aloud — direction belongs in voice_instruction, which is never spoken.
- Mixing @Audio references with a character image. They are mutually exclusive — pick voice cloning (up to 3 clips) or one image per prompt, not both.
- Overcrowding a scene. Ten characters in a two-minute generation blur together; two to four distinct voices per scene stays clean.
- Regenerating instead of Extending. A fresh generation recasts the voices; Extend continues the exact same voices, music and mix.
- Forgetting "loopable" for ambience — and describing one-off events (a scream, a crash) inside something meant to repeat.
- Changing voice or instruction mid-project. For audiobooks and series, save the exact voice + voice_instruction pair and reuse it verbatim every chapter.