When was Seed-Audio 1.0 released?

Seed-Audio 1.0 (Chinese name: 豆包音频生成模型 1.0) was announced on June 23, 2026 at Volcano Engine's FORCE conference in Beijing.

Is Seed-Audio 1.0 open source?

No. As of July 2026 there are no open weights and no technical paper. Access is API-only through Volcano Engine, BytePlus, and licensed third-party hosts such as fal.ai.

How is Seed-Audio different from Seed-TTS?

Seed-TTS (2024) is a text-to-speech family that renders a single voice. Seed-Audio 1.0 generates a complete audio scene in one pass: multiple consistent voices plus music, sound effects and ambience, already mixed.

Model Guide

Seed-Audio 1.0, explained without the hype.

Seed-Audio 1.0 is an audio scene generation model from ByteDance’s Seed team, announced June 23, 2026 at the Volcano Engine FORCE conference. It is not another TTS voice. One prompt returns finished audio: multi-character dialogue, music, sound effects and room tone, mixed in a single pass.

Try it in the playground

The specs that actually matter

ModalityText to full audio scene: dialogue, music, SFX, ambience in one generation

Max length2 minutes per take (long-form via continuation)

Prompt sizeUp to 2,048 characters on most hosts; voiced text reads best under 400 characters per take

Voice cloningZero-shot from up to 3 reference clips (30s each), referenced inline as @Audio1 to @Audio3

Image guidanceOne reference image can steer the scene instead of audio references

LanguagesEnglish and Chinese, including regional accents

Outputwav, mp3, pcm, ogg_opus at 8 to 48kHz

StreamingNot supported. Non-streaming only; ElevenLabs and MiniMax win for realtime use

Open weightsNone. No paper, no published benchmarks as of July 2026

Every way to access it

There is no self-serve consumer product from ByteDance yet, which is why third-party playgrounds exist. The real routes, verified in July 2026:

BytePlusFirst-party international API. $0.15 per minute of output, 300 free minutes on activation. Requires business verification

Volcano EngineFirst-party China API (invited beta) plus a 30-minute personal trial in the Ark experience center. China account required

fal.aiLicensed partner host, self-serve. $0.1875 per minute of output

AggregatorsVarious relays undercut official pricing. Supply legitimacy and stability vary; treat as unverified

YifThis site: free web takes now, credit packs launching. We relay official APIs and say so

Full price math lives on the pricing page, including why some aggregator rates look impossibly cheap.

Seed-Audio vs the rest of the Seed family

ByteDance ships several audio-adjacent models and the names get mixed up constantly:

Seed-TTS (2024): single-voice text-to-speech research family. The “Seed-TTS eval” benchmark you see in papers belongs to this older work, not to Seed-Audio 1.0.
Seed-Music: music generation. Seed-Audio can include music in a scene but is not a dedicated music model.
Seed Realtime Voice: low-latency conversational voice for assistants. Streams, but does not build scenes.
Seed-Audio 1.0: the scene model. Dialogue plus music plus SFX plus ambience, one pass, pre-mixed.

What it is genuinely good at

Multi-character dialogue where every voice stays consistent across the whole take
Scenes that need a soundtrack and effects placed with intent, not layered afterwards
Audio dramas, podcast intros, game barks, ad spots, guided meditations

Where it loses today

Realtime and streaming use cases: not supported at all
Very long single takes: 2 minutes per generation, continuation required beyond that
Language coverage beyond English and Chinese

Want the grammar that makes it sing? Read how to prompt Seed-Audio 1.0, then render your first scene.