Seed Audio 1.0 is ByteDance's text-to-audio model — give it a script and it speaks it, or give it a scene direction ("a late-night suspense radio drama") and it produces the whole thing: voices, pacing, atmosphere. 20 preset voices, voice cloning from reference clips, and delivery controls for speed, volume, and pitch. 0.1 credits per second of output.
Seed Audio 1.0 works in two modes, and telling them apart is the key to prompting it well. Verbatim mode: give it the exact words to speak (wrap the script in quotes) and it reads them in the voice you pick. Instruction mode: describe what you want ("generate an energetic podcast intro for a show about AI tools") and the model writes and performs the whole piece — script, delivery, even production feel.
Voices come three ways. Pick one of 20 presets by name (kian_en_zh, vivi_mixed_en_zh_ja_es_id, monkey_king_zh…), leave the voice unset and let the model cast the right one for your prompt, or clone a voice by passing up to three reference clips through the API and pointing at them in the prompt as @Audio1–@Audio3.
Billing is per second of finished audio: 0.1 credits per second. A hold is charged when you submit based on the prompt length, then settles to the exact output duration when the clip finishes — you only ever pay for the audio you get.
20
Preset voices
0.1 cr
Per second of output
3
Reference clips for cloning
The headline capability: describe a scene instead of writing a script. Radio dramas, podcast intros, ad spots, guided meditations — the model writes the material, casts the voice, and performs it with fitting pacing and tone. Short instructions produce surprisingly complete pieces.
Named presets span English, Chinese, Spanish, Japanese, Indonesian, and Portuguese — from warm narrators (kian_en_zh) to multilingual all-rounders (vivi_mixed_en_zh_ja_es_id) to characters like monkey_king_zh. Leave voice unset and the model picks one that fits the prompt.
Pass up to three reference clips (30s / 10MB each, wav / mp3 / pcm / ogg) via audio_urls in the API, then reference them in the prompt as @Audio1, @Audio2, @Audio3. The model synthesizes in the referenced voice. A single reference image (image_url) can guide the vibe instead — one or the other, not both.
Speed runs 0.5x–2x, volume 0.5–2, pitch shifts ±12 semitones. Use speed for pacing (1.1x tightens ad reads), pitch for character work (-3 semitones deepens a villain), and leave volume at 1 unless you are matching existing audio.
Mixed-language presets switch languages mid-sentence without breaking the voice. Write the prompt with the languages inline — English, Spanish, and Chinese in one line reads naturally with a voice like vivi_mixed_en_zh_ja_es_id.
Output as mp3 (default, smallest), wav or pcm (uncompressed, for editing pipelines), or ogg_opus. Sample rates from 8 kHz (phone-line feel) to 48 kHz (production). 24 kHz default is right for most voice work.
Each example shows the exact prompt that produced the result. Copy any prompt with one click.
Instruction mode · voice auto-cast
Generate a short suspense radio drama set in a late-night convenience store: a nervous clerk, a mysterious customer, rain against the windows, tense music sting at the end.
Pure instruction mode — no script given. Name the genre ("suspense radio drama"), the setting, the cast, and one or two production beats ("rain against the windows", "tense music sting"). The model writes and performs the whole scene.
Verbatim mode · kian_en_zh
Read this in a warm, measured documentary narrator voice: "Every image begins as noise. In less than a second, a billion tiny decisions turn that noise into a photograph that never existed. This is the story of generative art."
Verbatim mode — the exact script sits in quotes, and the delivery direction ("warm, measured documentary narrator") sits outside them. Pairing the direction with a steady preset like kian_en_zh keeps the read consistent across takes.
Instruction mode · voice auto-cast
Generate an energetic morning-show podcast intro: an upbeat host welcoming listeners to "The Creative Stack", teasing today's episode about AI image tools, with a short jingle feel.
Instruction mode with a quoted proper noun. Show names, brand names, and taglines in quotes survive intact while the model improvises everything around them. "With a short jingle feel" steers the production style.
Verbatim mode · sophie_en_zh
Read this line as a mischievous fantasy villain, slow and theatrical: "You really thought a firewall could keep me out of the kingdom's archives?"
Character direction works like directing an actor: an archetype ("mischievous fantasy villain") plus two delivery adverbs ("slow and theatrical"). Add a pitch shift of -2 or -3 semitones for extra menace without changing the voice.
Verbatim mode · mindy · 1.1x speed
Read this as a bright 20-second ad spot: "PixelDojo puts seventy AI models in one studio. Images, video, and now audio. Start creating at pixeldojo.ai."
Ad reads want tempo: naming the format ("bright 20-second ad spot") sets the energy, and bumping speed to 1.1x gives it broadcast tightness. URLs and brand names read cleanly without special markup.
Verbatim mode · vivi_mixed_en_zh_ja_es_id
Read this naturally, switching languages mid-sentence: "Welcome to the future of sound. Bienvenido al futuro del sonido. 欢迎来到声音的未来。"
Code-switching needs a mixed-language preset — vivi handles English, Spanish, and Chinese in one breath with the same voice identity. Write the languages inline exactly as they should be spoken.
Text inside double quotes is spoken verbatim; everything outside is direction. No quotes at all reads as an instruction, and the model writes its own script. Mixing them ("Read this as X: '…'") is the most reliable pattern for controlled reads.
Two or three delivery words beat a paragraph of description: "warm and measured", "slow and theatrical", "bright and fast". Name an archetype (narrator, morning-show host, villain) to anchor the overall performance.
Leaving the voice unset is genuinely good — the model picks a voice that fits the material, which matters most in instruction mode where it also wrote the script. Lock a named preset when you need the same voice across takes.
Do not fix pacing in the prompt when a knob does it better: 1.1–1.2x for ad tightness, 0.9x for meditations. Pitch ±2–3 semitones changes character weight while keeping the voice recognizable; ±12 is a full octave and gets cartoonish fast.
At 0.1 credits per second, a 30-second clip costs 3 credits. The up-front hold estimates from your prompt length and settles to the real duration on completion — instruction prompts that produce long scenes settle up, over-estimates settle back.
| Setting | Values | Notes |
|---|---|---|
| prompt | up to 2,048 chars | Script (in quotes) or scene instruction. @Audio1–@Audio3 reference uploaded clips. |
| voice | preset name or cloned id | Omit to let the model cast. 20 presets across 6 languages. |
| audio_urls | up to 3 URLs (API) | 30s / 10MB each, wav / mp3 / pcm / ogg_opus. Not combinable with image_url. |
| image_url | 1 URL (API) | jpeg / png / webp up to 10MB. Alternative to audio references. |
| output_format | mp3 · wav · pcm · ogg_opus | mp3 default. wav/pcm for editing pipelines. |
| sample_rate | 8k–48k Hz | 24,000 Hz default — right for most voice work. |
| speed | 0.5–2.0 | 1.0 normal. 1.1x tightens ad reads. |
| volume | 0.5–2.0 | Leave at 1 unless matching existing audio. |
| pitch | -12 to +12 semitones | ±2–3 for character weight; ±12 is a full octave. |
| Cost | 0.1 credits / second | Held from a prompt-length estimate, settled to the exact output duration. |
You are charged 0.1 credits per second of finished audio. Submitting holds an estimate based on your prompt length; when the clip completes, the charge settles to the exact duration — longer outputs collect the difference, shorter ones refund it automatically.