Seed Audio 1.0 by ByteDance Seed: An Honest Explainer (Text-to-Audio, Not Just TTS)

Every page about Seed Audio 1.0 recycles the same vendor bullet list. This one does three things those pages skip: it separates API-doc-verified specs from marketing, normalizes the scattered reseller prices into a like-for-like table, and states plainly that no independent benchmark exists. That honesty is the whole point.

What Seed Audio 1.0 actually is

Seed Audio 1.0 is ByteDance Seed's all-in-one text-to-audio model. It was announced on June 23, 2026 at Volcano Engine FORCE 2026; some API docs list a June 29, 2026 availability date, which likely reflects the API go-live rather than the event announcement. The core idea, consistent across every source we checked, is that it renders a full sound scene from a single prompt: spoken dialogue, background music, ambience, and foley-style sound effects, all composed in one generation pass.

That framing matters. Most tools you have heard of are text-to-speech only. They turn words into a voice. Seed Audio is positioned as text-to-any-audio: describe a scene and it produces the whole soundstage at once. That single-pass composition is the honest differentiator, and it is corroborated by three independent API-provider docs that actually host the model: fal.ai, Segmind, and WaveSpeed.

Before we go further, a disclosure that no competing page seems to make: we did not benchmark this model ourselves. This is an explainer, not a first-party test. Where a claim is verified against rendered API documentation we say so; where a claim is vendor or reseller marketing we flag it; and where evidence simply does not exist, we say that too.

Verified specs vs marketing claims

Here is the separation the SEO pages never draw. The left column below is what two or three independent provider API docs agree on. The right column is what appears only in vendor or reseller marketing and is not confirmed in the rendered first-party docs.

A note on the audio reference rule that matters if you plan to build with it: per the fal.ai and Segmind docs, audio references cannot be combined with an image input. If your workflow needs both, plan around it.

How it differs from TTS-only tools

The lazy version of this comparison says "ElevenLabs can't do X." That is not true and we will not write it. ElevenLabs has added multi-speaker dialogue and sound-effects features too. The fair, defensible difference is architecture and access: Seed Audio generates voice, music and SFX together in a single pass, and it is delivered primarily through ByteDance's enterprise clouds. That is a genuine distinction in the model and the access model, not a capability gap you can wave around.

If you are mapping the broader generative-media landscape, our companions on best text-to-video 2026 and best text-to-image 2026 cover the visual side of the same shift toward single-prompt scene generation.

Normalized reseller pricing

This is the part every other page gets wrong or omits. There is no confirmed first-party Volcano Engine list price. Every number below is a third-party reseller rate, as of research date, and they differ. Do not present any of these as ByteDance's official rate.

The clean takeaway: on a per-minute basis the verified rates cluster loosely around $0.12–$0.20 per minute, with fal.ai's $0.1875/min as the most defensible anchor because it is a flat, rendered figure. WaveSpeed and Atlas Cloud price on different units (per-run, per-token), so they are not directly comparable and we do not force them into a false equivalence.

What a real 2-minute scene costs

Numbers in isolation do not help you budget. So here is the concrete reframe no competing page offers. Suppose you generate a 2-minute multi-character scene — the model's stated per-generation ceiling — at the cleanest verified anchor of $0.1875/min:

2 min × $0.1875 = $0.375 per finished scene (fal.ai, modeled from the verified per-minute rate).

At the low end of the aggregated spread ($0.002/sec = $0.12/min) the same scene is about $0.24; at the high end ($0.00325/sec = $0.195/min) it is about $0.39. So a full two-minute dialogue-plus-music-plus-SFX scene lands roughly in the $0.24–$0.39 band depending on provider. These are modeled figures derived from published reseller per-second/per-minute rates, not a first-party quote and not a bill we ran.

This is the same cost-per-task lens we apply to visual models in Sora 2 cost and Nano Banana pricing: don't compare unit prices, compare what one finished deliverable actually costs.

How to access it

There are three official channels and several third-party APIs. Be careful here: many top search results (seedaudio.im, seedaudio.co, seedaudio.run and similar) are affiliate/SEO domains, not ByteDance. The only official channels are below.

• Volcano Engine — ByteDance's China enterprise cloud, via API. This is the enterprise access reality.
• BytePlus — for international developers; an official doc exists at docs.byteplus.com (the product is listed as "Audio 1.0 / Seed Speech"). The page is a single-page app that did not fully render for us, so treat the listing as confirmed-to-exist rather than fully read.
• Doubao app — consumer access. Integrations into CapCut/Jianying, Jimeng and Fanqie have been reported, but only in secondary sources.
• Third-party APIs — fal.ai, WaveSpeed, Atlas Cloud and Segmind host the model at the differing prices covered above.

If the "gateway" model of access is new to you, our primer on what an LLM gateway is explains why routing many models through one endpoint beats juggling a dozen provider keys — the same logic that makes multi-provider audio access manageable.

Honest verdict

Seed Audio 1.0 is real, hosted on multiple live APIs, and its core capability — one prompt to a full sound scene in a single pass — is well corroborated. If you want dialogue, music and effects composed together and you are comfortable with ByteDance-cloud access, it is worth a trial.

But be clear-eyed about what is not established. There is no technical paper, no independent benchmark, and no first-party press coverage. Quality claims — foley realism, "better than SeeDance 2" — trace to vendor demos or a single social-media anecdote (an X post from @TomLikesRobots claiming a ~15-second test cost a couple of cents). The six-language preset list is marketing, not confirmed in rendered docs. And no official price has been published. Treat the enthusiasm accordingly, run your own listening test on a real script, and price it per finished scene, not per unit.

FAQ