T2A, TA2A, references, voices, music, ambience, and effects
This guide turns the prompt-writing notes into a practical page for creators: how to structure text-only prompts, how to map audio references, and how to describe voices, BGM, ambience, and sound effects without overloading the model.
T2A
Text or image guided audio
TA2A
Text plus up to 3 audio references
1 image
Image and audio references should not be mixed
<= 30s refs
Keep each reference clip short and focused
Short version
Seed Audio 1.0 can combine speech, character delivery, ambience, music, and foley-style events. The safest way to guide that mix is to write a short production brief in playback order: first the scene, then the speakers, then the sound layers, then the ending.
Templates
Create a [duration] [format] in [language].
Scene: [place, time, mood, audience].
Voice: [speaker count, speaker traits, emotion, pace, clarity].
Dialogue or narration: "[exact words or conversation outline]".
Ambience: [room tone, weather, crowd, traffic, environment].
Music: [style, instruments, intensity, entry or fade].
Sound effects in order: [event 1], [event 2], [ending cue].
Keep the result [clean / cinematic / minimal / broadcast-ready].Use @audio1 for [speaker or voice trait].
Use @audio2 for [second speaker or style reference].
Create a [duration] [format] in [language].
Speaker map: [speaker A follows @audio1], [speaker B follows @audio2].
Scene: [place, mood, purpose].
Delivery: [emotion, rhythm, accent, pauses, laughter, hesitation].
Dialogue: "[lines or conversation structure]".
Layers: [ambience], [music bed], [important sound effects in order].
Keep voices consistent and keep speech easy to understand.Prompt formula
Name the format, setting, audience, language, mood, and rough length before writing dialogue.
Define speaker count, speaker traits, vocal texture, accent, emotion, pacing, and clarity priority.
Write the prompt in playback order so speech, ambience, music, and sound events do not compete.
Add one ambience bed, one music direction, and only the sound effects that matter to the scene.
Input modes
The content can look similar, but the job is different. T2A asks the model to invent a scene. TA2A asks it to follow reference signals while still obeying the written direction.
Use this when you want Seed Audio 1.0 to invent the voice and scene from a written brief, optionally guided by one image reference.
Use this when voice identity, pacing, emotional texture, or multi-speaker consistency matters. Map each reference inside the prompt.
Voice checklist
Layer timeline
Time words help the model understand priority. Use them to stage the mix, not to micromanage every frame.
0-5s
Room tone, weather, phone buzz, door open, or a simple music entrance.
5-25s
Dialogue or narration with the most important voice and emotion directions.
25-40s
One or two sound events that support the story without covering the voice.
Final beat
A clean ending cue, fade, resolved chord, bell, door, breath, or button sound.
Example prompts
These are rewritten examples, not copies from the source document. Use them as structure, then swap in your own scene, speakers, and sound events.
Create a 45-second cinematic crime scene in English. Start with a quiet phone vibration, distant rain, and a low suspense pad. Detective A speaks in a restrained, tired male voice, close to the microphone. Detective B answers through a slightly distorted phone line with a calm but evasive tone. Keep the dialogue clear. Add two small footsteps and one car brake near the end.Use @audio1 for Host A and @audio2 for Host B. Create a 60-second energetic livestream shopping segment in English. Host A is bright, fast, and persuasive. Host B is warmer and reacts with short supportive lines. Add very light upbeat background music, subtle package handling sounds, and a clean call-to-action ending.Use @image as the character reference. Create a 35-second fantasy narration in English. The speaker is a serious middle-aged astrologer with a slow, ceremonial delivery and a deep resonant voice. Add quiet rain outside a stone observatory, a soft page turn, and a distant bell. Keep the music minimal and mysterious.Common mistakes
A prompt with five music moods, ten sound effects, and several character arcs usually becomes muddy. Reduce the mix first.
For TA2A, do not just upload files. Say which character or style should follow @audio1, @audio2, or @audio3.
Seed Audio 1.0 works better when the prompt reads like a compact production brief in playback order.
Extreme emotional adjectives can make delivery less natural. Start with a controlled cue, then intensify only one detail.
Requested length is guidance. Leave room for trimming when the final edit needs frame-accurate timing.
Describe roles and vocal qualities generically instead of naming real people, celebrities, public figures, or protected characters.
Start with the templates here, then compare them against playable outputs in the tested how-to guide.