Seed Audio 1.0: The Essentials
Last updated: July 7, 2026
Seed Audio 1.0 by BytePlus is a text-to-audio performance model. From a single text prompt it renders a full scene: one or more characters speaking, background music, and sound effects, all in one pass. It is not a traditional text-to-speech engine with a fixed voice list. You describe the voice and the scene in plain English, and the model performs it.
Ask for a sweet 22-year-old with a soft French accent, a gravelly 70-year-old grandfather telling a bedtime story, or a full battlefield with three named heroes over war drums and clashing steel. The age, accent, mood, characters, music, and effects all live inside the prompt.
Overview
Most speech models take a line of text and read it in a preset voice. Seed Audio 1.0 works differently: the prompt describes the performance, and the model casts the voice, sets the emotion, and layers in the score and the sound design around it. That makes it closer to a voice director than a narrator.
It runs in four ways, and you pick the one that fits the job:
Text to Audio (T2A): describe everything in text (voices, music, SFX, dialogue) and generate the whole scene from words alone.
Text plus Audio to Audio (TA2A): attach up to 3 reference clips and point to them in the text as @Audio1, @Audio2, @Audio3 to lock specific character voices while still describing the scene.
Voice cloning: attach a single clean reference clip and have the model speak new text in that voice.
Voice from image: attach one reference image and let the model infer a voice from the character it shows.
What It Does
Casts a described voice: age, gender, accent, timbre, mood, and archetype, all from the prompt.
Performs multi-character dialogue with distinct voices in a single generation.
Adds background music and sound effects in the same pass, so a scene arrives already mixed.
Handles narration, ad reads, NPC crowds, cinematic scenes, and pure music-and-SFX beds.
Speaks many languages naturally, including strong Brazilian Portuguese and clean code-switching between languages inside one line.
Gives fine control over speed, loudness, and pitch when the description alone is not enough.
How to Use Seed Audio 1.0
Open the model page. For a simple single-voice line, write a prompt with two parts:
The voice description. Lead with the character: age, gender, accent, tone, mood, archetype. Be concrete. "Bright, slightly high-pitched voice of an anime hero in his late teens, full of conviction" performs better than "young excited boy".
The line to speak. Quote the exact text you want delivered. The model speaks what is inside the quotes and treats the rest as direction.
A working single-voice template:
In the [warm, gentle, husky, gravelly] voice of a [age]-year-old [gender] with a [accent], read the following: '[the actual line to speak].'
For a full scene with music, effects, and several characters, BytePlus recommends a five-part structure. Cover these in order, and put a line break between each part so the prompt stays readable and easy to steer:
Environment: location, time, weather, space, and atmosphere.
Music and SFX: the music genre, tempo, and mood, plus the sound effects you want (war drums, metal scrapes, wind, crowd).
Character setup: for each character, the age, identity, personality, and voice.
Voice description: timbre, accent, speed, emotion, and qualities such as raspy, bright, breathless, or calm.
Dialogue: the exact lines each character speaks, with emotional and action cues, each line in quotes.
Three complete prompts, together with the audio they produced, are in the Examples section below: a five-part dialogue scene (example 2), a music-and-effects-only bed (example 6), and a reference-audio scene (example 12).
If you want tighter timbre control, attach an audio reference clip (up to 3, 30 seconds each) or a single image reference, and the model bases the voice on the reference instead of inferring it purely from the description. In the text, point to each clip as @Audio1, @Audio2, or @Audio3.
Parameters
Text
The full prompt: voice and scene description followed by the lines to speak. Up to 3000 characters. When using audio references, point to each one inside the text as @Audio1, @Audio2, or @Audio3.
Audio References
Up to 3 audio samples that seed the timbre. Each clip is capped at 30 seconds and 10 MB. Cannot be combined with an image reference. Useful when you have a target voice in mind that is hard to describe in words, or when a scene needs several clearly different voices.
Image Reference
A single reference image (JPEG, PNG, or WebP, up to 10 MB) that suggests a voice. Cannot be combined with audio references. Useful for character portraits where the look implies the sound.
Sample Rate
Audio quality, in samples per second. Choose from 8000, 16000, 24000, 32000, 44100, or 48000 Hz. Default 24000 is a good balance of clarity and file size for speech. Use 44100 or 48000 when the output carries music and effects, or will be mastered alongside other audio.
Speech Rate
How fast the voice speaks. Range: -50 (half speed) to 100 (double speed). Default 0. Keep within plus or minus 20 for natural delivery; extreme values stretch or compress audibly.
Loudness
Voice volume. Range: -50 (half) to 100 (double). Default 0. Use small positive values to lift a voice above a loud music bed, or small negatives for a whispered, intimate delivery.
Pitch
Voice pitch in semitones. Range: -12 to 12. Default 0. Subtle shifts (plus or minus 2) can age a voice or change its character without breaking realism. Larger shifts cross into cartoon territory.
Examples
Twelve scenes generated on Scenario, each from a single prompt. They cover multi-character dialogue, cinematic narration, pure music and effects, reference-locked voices, and single-voice reads across games, marketing, film, education, and events. Press play to listen, or open each one on Scenario. Three of them (2, 6, and 12) include the complete prompt used, exactly as submitted.
1. Battlefield scene, three heroes over a war score (games)
Full T2A scene: an environment, an orchestral score, sound effects, and three named characters (Kael, Sena, Doran), all in one prompt. Settings: sample rate 48000, loudness +5.
Listen (MP3) · Open on Scenario
2. Temple dialogue, a complete five-part prompt (games)
Four characters defined up front, then scene beats and quoted lines, with a line break between every part. Adapted from BytePlus's own showcase and trimmed to fit the 3000 character limit. Settings: defaults.
Listen (MP3) · Open on Scenario
The complete prompt, exactly as submitted:
Background music: gentle harp arpeggios, soft wooden flute, warm strings, light hand drums, and distant children's choir humming without words. The music should feel adventurous, pure, ancient, and hopeful, with a slight sense of mystery.
Sound effects: morning birds, leaves moving in a soft breeze, glowing magic particles, stone mechanisms shifting, distant waterfall, footsteps on moss-covered stone, a small creature's wings fluttering, and faint temple chimes.
Characters:
Arin: a young male adventurer, brave but inexperienced, warm youthful voice, slightly breathless from climbing, curious and sincere.
Mira: a small glowing forest spirit, high soft voice, playful, fast-talking, magical, sometimes teasing but deeply loyal.
Elder Rowan: an ancient tree guardian, very deep slow voice, resonant, wise, calm, with a wooden texture, as if the forest itself is speaking.
Kael: a masked rival warrior, young adult male voice, cool, sharp, controlled, proud but not evil, hiding concern beneath sarcasm.
[Scene opening]
Background music: soft harp and flute begin gently. A distant waterfall flows behind the temple. Tiny magical particles shimmer in the air.
Sound effects: leaves rustle; birds call; Arin steps onto old stone; moss squishes softly under his boots.
Arin, young male voice, breathless and amazed, whispering with wonder, says: "So this is it… the temple from the old map."
Mira, tiny glowing spirit voice, playful and excited, fluttering around him quickly, says: "Old map? Please. That thing was half-burned, upside down, and probably drawn by a goat."
Arin, amused but still focused, with a quiet laugh, says: "And yet it brought us here."
Mira, teasing but fond, with a bright magical giggle, says: "Fine, brave hero. I give the goat full credit."
Sound effects: a small stone tile sinks under Arin's boot with a soft click. Ancient gears begin moving beneath the floor.
Arin, suddenly tense, voice dropping into alarm, says: "Wait… did you hear that?"3. Fantasy party banter, three distinct voices (games)
A knight, a rogue, and a mage in a dungeon corridor, each with a clearly different voice and personality. Settings: defaults.
Listen (MP3) · Open on Scenario
4. Cinematic Norse narration with score and ambience (film)
A single low, ominous narrator over a dark score, wind, and distant bells. Slowed and pitched down for gravity. Settings: sample rate 44100, speech rate -15, pitch -2.
Listen (MP3) · Open on Scenario
5. Tavern crowd with a quest hint (games)
Overlapping NPC voices, a crackling fireplace, and a bard, with one key quest line kept slightly clearer than the rest. Settings: defaults.
Listen (MP3) · Open on Scenario
6. Boss-fight underscore, music and effects only (games)
No voice at all. A dark heroic orchestral bed with taiko drums, brass, and combat sound effects, driven purely by the prompt. Settings: sample rate 48000, loudness +8.
Listen (MP3) · Open on Scenario
The complete prompt. Note the opening instruction that excludes voice, and the timestamped structure:
Generate pure background music and sound effects only. No voice, no dialogue, no narration, no lyrics. Style: epic final boss battle, dark heroic orchestral with hybrid percussion. The mood should feel massive, dangerous, and adrenaline-driven. Create original music only. Background music: thunderous taiko and frame drums, aggressive low brass, fast tremolo strings, a dark ominous choir, and a soaring heroic string theme in the climax. Sound effects: a giant creature roar, heavy stomps shaking the ground, sword clashes, magic charge-ups, fireball impacts, shield blocks, and a stone arena crumbling. Structure: 0:00-0:10 boss reveal roar with a huge brass hit; 0:10-0:30 relentless drum-and-brass assault with combat SFX; 0:30-0:45 tense mid-fight with tremolo strings and magic charges; 0:45-0:60 heroic climax with full choir, then a final crushing impact. Final mix: no vocals or words, original, cinematic, and powerful.7. Product launch voice-over (marketing)
An energetic single announcer over an upbeat bed, pushed faster for hype. Short and punchy. Settings: sample rate 44100, speech rate +20, loudness +5.
Listen (MP3) · Open on Scenario
8. Horror intro, close whisper (film)
A trembling whisper over a dissonant drone and a warped music box. Slowed, pitched down, and quieted for dread. Settings: sample rate 44100, speech rate -20, pitch -3, loudness -8.
Listen (MP3) · Open on Scenario
9. Kids adventure duo, bright and playful (games)
A young hero and a squeaky cloud spirit trade fast, funny lines over cheerful ukulele, two very different voices. Settings: defaults.
Listen (MP3) · Open on Scenario
10. Bilingual event host, Portuguese and English (events)
A warm host welcomes a crowd in Brazilian Portuguese, switches to English, then back, in a single natural take. Settings: defaults.
Listen (MP3) · Open on Scenario
11. Science explainer, friendly teacher (education)
A warm, clear instructor voice walks through why the sky turns red at sunset, with soft cues between ideas. Settings: defaults.
Listen (MP3) · Open on Scenario
12. Locked voices with reference audio, TA2A (film)
An interrogation scene where each character's timbre is pinned to a reference clip instead of a written description. The prompt points to the clips as @Audio1 (the detective) and @Audio2 (the suspect); both reference voices were themselves generated with Seed Audio 1.0 from short text descriptions. Settings: sample rate 48000.
Listen (MP3) · Open on Scenario
The complete prompt. Voice descriptions stay in the text as direction, and each speaker tag names its reference clip:
A dim, rain-streaked interrogation room. A single flickering fluorescent light hums overhead. Distant thunder rolls outside.
Sound effects: rain tapping against a window, a metal chair scraping against concrete, the low hum of the fluorescent light.
Detective Hale (a warm, mature male voice, tired but controlled, voiced by @Audio1) says quietly: "You've been lying to me since you walked in here."
Mara (a bright but now shaken young female voice, trying to stay composed, voiced by @Audio2) answers, defensive: "I told you everything I know."
Sound effects: the detective sets a folder down on the table with a soft thud.
Detective Hale (voiced by @Audio1), calm but pointed: "Then explain why your car was on that street at midnight."
Mara (voiced by @Audio2), voice cracking slightly, quieter now: "Because, I was scared. I didn't know who else to call."
Sound effects: a long pause; the rain intensifies against the window.
Detective Hale (voiced by @Audio1), leaning back, voice softening just slightly: "Then start from the beginning. And this time, don't leave anything out."Tips for Better Results
Lead with the character, not the line. The first sentence of your prompt should describe the voice. The line to speak goes in quotes after the description. Reversing the order weakens the casting.
Use the five-part structure for scenes. Environment, then music and SFX, then character setup, then voice description, then dialogue. It gives the model everything it needs to mix a scene rather than just read a line.
Format the prompt with line breaks. Put each character and each section on its own line. A long scene written as one unbroken block is hard for you to read and harder to steer.
Quote every spoken line literally. The model treats quoted text as the script and unquoted text as direction. Label each line with the character and an emotion cue.
Be specific about age and tone. "Late teens, bright and passionate" beats "young excited". "Late forties, husky and tired" beats "older woman".
Name the region and the language for accents. "English with a soft French accent" works. "French accent" alone is more ambiguous. Portuguese and other languages perform well, so name them directly.
Lock important voices with a reference. If a scene depends on several clearly different characters, attach up to 3 reference clips and point to them with @Audio1, @Audio2, @Audio3 rather than relying on descriptions alone.
Leave the dials at default first. The description usually gives you the pacing and pitch you want. Reach for speech rate, loudness, and pitch only to fine-tune, and keep music-heavy scenes at 44100 or 48000.
State the target length for short assets. For stingers, hooks, and UI-length clips, open with "Total duration: exactly 15 seconds" and timestamp each beat, such as [0:00-0:04] for the intro and [0:04-0:09] for the next cue. In testing, short scenes written this way landed on the requested length, while long, dense scripts drifted well past it.
Retry before rewriting. The content check screens the generated audio, not just your text, so a perfectly benign prompt can fail once and pass unchanged on the next run. If a run fails and the prompt is clearly harmless, generate again before touching the text.
Simplify punctuation if failures persist. Repeated failures on the same prompt sometimes clear up after replacing typographic marks such as em dashes and ellipses with commas, colons, or periods. Plain punctuation gives the parser less to trip on.
Known Limitations
3000 character limit on the prompt. BytePlus's own showcase prompts often run longer than this. When adapting one, trim at a clean sentence or quote boundary rather than cutting mid-line, so the scene still reads cleanly.
Duration follows the text and can drift. Output length tracks the amount of text. Simple prompts lock to a clean, predictable length, but long or complex multi-character scenes can run long and occasionally cut off a final line. Keep the most important line early, and split very long scenes into parts. For short assets, an explicit total duration with timestamped beats (see the tips above) holds the output on target.
Distinct characters can drift toward one timbre. Two described characters sometimes come out sounding similar in the same generation. When clearly different voices matter, lock each with an audio reference.
Voice consistency across runs. The same description does not guarantee the same voice every time. For a recurring character, capture a good output and reuse it as a reference clip.
Audio and image reference are mutually exclusive. Use one or the other in a single run, not both.
Reference clips are short. 30 seconds and 10 MB each, up to 3 total. Trim long source recordings before upload.
Extreme pitch and speed degrade naturalness. Push speech rate past plus or minus 30, or pitch past plus or minus 6 semitones, and the voice starts to sound processed.
Occasional false positives from the content filter. The filter reviews the generated audio after the run, and it sometimes flags a harmless generation with a content policy or formatting error. The same prompt often succeeds on the next attempt, so retry once or twice before rewriting, and simplify unusual punctuation if the failures continue.
Use Cases
Games: cinematic narration and cutscenes, multi-character dialogue and battle banter, NPC crowd chatter with quest hints, and music-and-SFX beds, all from prompts.
Marketing and advertising: ad voice-over with a specific character, hype spots with a music bed, brand persona exploration before casting real talent.
Film and animation pre-vis: scratch tracks and animatic scenes with score and effects, character voice tests before recording sessions.
Education and e-learning: friendly instructor voices, language-learning samples with regional accents, explainer narration.
Localization and events: multilingual reads and bilingual hosting, and code-switching in a single take.
From Prompt to Pipeline
For teams putting Seed Audio 1.0 into a real production loop, BytePlus recommends a simple five-step workflow:
Build a character voice library. Create reference clips and a short written voice spec for main characters, villains, and key NPCs. You can bootstrap the clips with Seed Audio itself: describe the voice, generate a clean line, and keep the best take as the reference.
Template your prompts. Write reusable prompt templates per scene type: battle, story beat, exploration, shopkeeper dialogue, narration, skill callouts. Swap in the lines and keep the structure.
Generate multiple candidates. For important lines, produce several versions and pick the best. The same prompt casts a slightly different voice each run, which works in your favor here.
Review before shipping. Check line accuracy, emotion, pronunciation, character consistency, volume, and unwanted sounds.
Wire the assets into your engine or your edit. Trigger the approved clips from story nodes, battle events, distance, or state machines, or cut them into your video timeline.