Seed Audio 1.0: What ByteDance's Audio Scene Model Actually Does

Two weeks after its June 29, 2026 release, Seed Audio 1.0 is showing up in a scattered pattern: a BytePlus onboarding wiki, a handful of aggregator listings, per-second pricing on at least four API providers, and a small cluster of hands-on threads on X. No ByteDance press event, no arXiv paper, no benchmark table. Just a model that appears to be quietly shipping through partner channels while the community figures out what it actually does.

TLDR Seed Audio 1.0 is ByteDance's text-to-audio model that generates multi-character dialogue, ambient sound, and sound effects from a single prompt. It supports zero-shot voice cloning from up to three reference clips (plus one optional reference image), outputs up to 120 seconds per request, and is priced at $0.002–$0.00325 per output second across providers. Primary provider docs describe it as text-to-audio with reference inputs; some third-party writeups describe video-conditioned generation, which remains unverified.

Key Takeaways

• Seed Audio 1.0 was released June 29, 2026, per the AIMLAPI model page.
• Multi-character dialogue, background music, ambient sound, and effects generate in a single pass, according to the BytePlus onboarding wiki.
• Reference inputs are capped at three audio clips and one image per request.
• Per-second pricing spans $0.002 (PiAPI) to $0.00325 (AIMLAPI), with Fal.ai at ~$0.003125/sec.
• No official benchmarks or architecture paper have been published in the signal set reviewed.
• Whether video conditioning is a first-class input remains contested between vendor docs and third-party writeups.

What Was Actually Shipped

The concrete facts, drawn from primary sources and hands-on user reports:

• Release date: Seed Audio 1.0 was released on June 29, 2026, per the AIMLAPI listing.
• Advertised capability: Multi-character dialogue, natural emotion, background music, ambient sound, and sound effects in a single generation, per ByteDance's own Lark onboarding wiki.
• Reference inputs: Up to three reference audio clips and up to one reference image per request, per the Atlas Cloud API schema.
• Text limit: 2,048 Unicode characters per request, per the PiAPI documentation.
• Max output duration: 120 seconds per generation.
• Formats: WAV, MP3, PCM, OGG Opus; sample rates from 8,000 Hz to 48,000 Hz.
• Provider IDs: bytedance/seed-audio-1-0 on AIMLAPI, byteaudio with task_type: seed-audio-1.0 on PiAPI, bytedance/seed-audio-1.0 on Fal.ai and Atlas Cloud.
• Reference clip guidance: PiAPI's documentation recommends 5–10 seconds of clean reference audio for zero-shot cloning.
• Hands-on confirmation: X user @aditiitwt reported first-hand that one prompt generated multi-character dialogue with emotion, music, ambience, and effects together, matching the vendor's advertised behavior.

Why This Matters

The audio generation category has been fragmented for a long time. One model does text-to-speech. Another does music. A third does sound effects. Producing a finished 30-second radio drama has meant stitching outputs from three vendors and mixing by hand.

Seed Audio 1.0 is positioned as one of the first commercial models to compress that pipeline into a single generation. The Lark wiki calls it "the next-generation sound creation engine" and frames the multi-character-plus-ambience-plus-music output as a single-pass workflow. That framing is not new — Meta and Stability have gestured at it — but a shipped API with per-second pricing and reference-conditioning is a different thing from a research demo.

The second reason it matters is distribution. ByteDance has been steadily rolling out its Seed family — Seedance for video, Seed-TTS for speech, and now Seed Audio — through third-party API marketplaces before making noise directly. That pattern gives the wider developer community access without waiting for a regional rollout. It also produces the exact scattered-primary-source situation we see here: docs on Lark, listings on AIMLAPI, Fal.ai, PiAPI, RunComfy, and Atlas Cloud, and no unified vendor announcement page.

For teams already building on Seedance for video, a matched audio model from the same lab is worth watching. Whether the audio actually stays in sync with Seedance output is a separate question the bundle does not answer.

Advertised Capabilities

Pulling from the Lark wiki and the API schemas, the advertised capability surface breaks down as:

• Text-to-Audio: The default mode. A prompt produces speech and optionally music, ambience, and effects.
• Reference-Guided Generation: Up to three audio references or one image reference guide voice, style, and mood. References cannot be mixed — a single request uses either audio references or an image reference, not both, per the Atlas Cloud schema.
• Multi-Speaker Dialogue: The @Audio1, @Audio2, @Audio3 markers in the script assign different reference voices to different lines. This is what @aditiitwt described as "directing an entire audio scene with a single prompt".
• Voice Consistency: The Lark wiki calls out "rock-solid voice consistency" across long-form generations as a distinguishing feature — a claim that would matter for serial content like audiobooks, but one that has not been externally benchmarked.
• Speech Controls: speech_rate, pitch_rate, and loudness_rate are tunable at request time.
• Zero-Shot Cloning: Third-party guidance suggests 5–10 seconds of clean reference audio is sufficient. Seed Audio 1.0's advertised approach appears to be reference-based rather than fine-tuning-based.

The interesting design decision is the exclusivity constraint: an image reference cannot be combined with audio references in the same request. That points to two distinct conditioning pathways under the hood rather than a fused multimodal encoder.

Pricing Across Providers

Seed Audio 1.0's pricing is visible on four API providers, all charging per output second — a billing model that is unusual in text-to-speech (which usually charges per character) and closer to how video generation APIs bill.

Provider	Price	Unit	Notes
AIMLAPI	$0.00325	per output second	OpenAI-compatible SDK, model ID bytedance/seed-audio-1-0
PiAPI	$0.002	per output second	Async task API, max 120s output
Fal.ai	$0.1875	per output minute (~$0.003125/sec)	Handles reference audio and image input
Atlas Cloud	not disclosed in text excerpt	per output second	Standard async audio API
MindStudio	18 cents / minute ($0.003/sec)	per output minute	Third-party writeup, framed as flat rate

The spread — $0.002 to $0.00325 per second — is roughly 60% between the cheapest and most expensive listed vendor. For a 60-second scene, that is $0.12 vs $0.195. Not enormous in absolute terms, but material at scale.

For a rough anchor: at $0.002 per second, generating an hour of finished audio costs $7.20. A creator producing five hours of episodic content per week would spend around $36 per week on audio inference at the PiAPI rate. That is aggressive pricing for audio-scene output, though the comparison to per-character TTS pricing is not apples-to-apples because Seed Audio bundles multiple audio layers into one generated second.

Independent Provider Coverage

Beyond ByteDance's own BytePlus surface, Seed Audio 1.0 has landed on:

• AIMLAPI with an OpenAI-compatible endpoint.
• Fal.ai as a partner listing.
• PiAPI with an async task API and a Signal Studio playground.
• Atlas Cloud with a documented parameter schema including reference exclusivity rules.
• RunComfy as an API listing with limited public detail.
• A dedicated third-party portal at seedaudio.co selling monthly and yearly credit packs — not affiliated with ByteDance directly based on the page copy, but positioned as a Seed Audio-focused frontend.

That is a wider partner net than a soft launch usually gets. It reads as intentional distribution rather than a leak.

Community Signal

The X signal on Seed Audio is small but consistent. Three named accounts — @maxescu, @aditiitwt, @IamEmily2050 — describe the same underlying behavior: single-prompt generation of layered audio scenes with distinct character voices.

Alex Patrascu called it "The Seedance moment for audio" on the day of release. That framing — Seedance being ByteDance's video model that was itself a category-shifting release — is a strong analogy but a subjective one. There is no benchmark to back it in the signal set.

Emily reported using Seed Audio non-stop since day one for dialog and drama generation. This is a hands-on adoption signal but not a measured one.

Notably absent: no one in the signal bundle has posted a A/B comparison against ElevenLabs V3, Suno, or Seed-TTS. No one has published a WER (word error rate), no one has run a MOS (mean opinion score) study, no one has shared side-by-side clips of the same script rendered on Seed Audio and a competitor. The claims about quality are, so far, entirely qualitative.

Seed Audio 1.0 vs ElevenLabs V3: What the Signal Says

ElevenLabs V3 is the closest competitor named in the sources reviewed — implicitly, since almost every writeup on Seed Audio 1.0 defines it against "traditional TTS", which for most builders means ElevenLabs. Here is what the bundle actually supports:

• Scene generation vs voice generation: Seed Audio 1.0 is advertised as generating dialogue, music, ambience, and effects in one pass. ElevenLabs V3 is a text-to-dialogue model focused on voice quality and expressiveness. On this dimension, they are targeting different product surfaces — Seed Audio is going after audio scene generation; ElevenLabs is going after best-in-class voice.
• Reference cloning: Both support zero-shot cloning from short reference clips. Seed Audio's exact minimum is not stated in the primary docs; third-party guidance suggests 5–10 seconds. ElevenLabs V3's cloning behavior is not in this signal set for comparison.
• Max output length: Seed Audio 1.0 is capped at 120 seconds per generation via PiAPI. ElevenLabs V3's cap is not in this signal set.
• Pricing model: Seed Audio bills per output second ($0.002–$0.00325). ElevenLabs typically bills per character. Direct comparison is unverified — no public number from either lab in this signal set that normalizes the two billing models on the same audio output.
• Benchmark performance: Unverified — no public number from either lab in this signal set.

The honest read: Seed Audio 1.0 and ElevenLabs V3 do not directly overlap on the primary use case. Seed Audio is closer to a category of one — audio scene generation with dialogue baked in — than it is to a head-to-head TTS competitor. Whether the voice quality alone rivals ElevenLabs is exactly the comparison the community has not yet run publicly.

What We Know vs. What We Don't

What we know:

• Seed Audio 1.0 was released on June 29, 2026, according to the AIMLAPI model page.
• Seed Audio 1.0 can generate multi-character dialogue, natural emotions, background music, ambient sound, and sound effects from a single prompt, according to the BytePlus onboarding wiki and hands-on reports from @aditiitwt.
• Seed Audio 1.0 accepts up to three reference audio clips plus up to one reference image per request, per the Atlas Cloud and PiAPI API schemas.
• Seed Audio 1.0 is described as supporting English and Chinese, with broader language support planned, according to a third-party writeup from AI Adoption Agency.
• Seed Audio 1.0 has a maximum output duration of 120 seconds per request through PiAPI, with a text input limit of 2,048 characters.
• Seed Audio 1.0 pricing varies by provider: AIMLAPI lists $0.00325 per second of output, PiAPI lists $0.002 per second, and Fal.ai lists $0.1875 per minute (roughly $0.003125 per second).
• Seed Audio 1.0 supports zero-shot voice cloning from short reference clips, with third-party guidance recommending 5–10 seconds of clean reference audio for best results.
• Seed Audio 1.0 outputs WAV, MP3, PCM, and OGG Opus formats, with configurable sample rates from 8,000 Hz to 48,000 Hz, according to the Atlas Cloud and PiAPI schemas.

What we don't know:

• No official benchmarks have been published for Seed Audio 1.0 in the sources reviewed — vendor documentation focuses on capability descriptions rather than numerical evaluations.
• Coverage of video conditioning is inconsistent: one third-party writeup from MindStudio describes video-conditioned scene generation, while primary provider docs from PiAPI and Atlas Cloud describe it as text-to-speech with audio and image references. Video conditioning remains unverified from primary sources.
• The architecture of Seed Audio 1.0 has not been officially documented. A third-party writeup from Qwen3TTS speculates it uses latent diffusion, but no ByteDance paper or technical report has been cited in the sources reviewed.
• No independent comparison against ElevenLabs V3, Suno, Seed-TTS, or any other named competitor has been published in the signal set — quality claims remain qualitative.
• The full list of supported languages beyond English and Chinese is not disclosed in primary sources.
• Commercial usage terms vary by third-party provider and are not consistently documented in the sources reviewed.

How to Evaluate It Yourself

For teams considering Seed Audio 1.0 for production, a few pragmatic evaluation steps:

Generate a matched-script test set. Take five representative scripts from your actual content — a two-character dialogue, a monologue with emotion range, a scene with ambient sound requirements, a short-form ad, a long-form narration. Run each through Seed Audio 1.0 and your current audio pipeline. This is a more useful signal than any published benchmark.

Test the reference-clone consistency claim. The Lark wiki emphasizes "rock-solid voice consistency" across long-form generations. Generate ten separate 60-second segments using the same reference clip and audit whether the voice drifts. Long-form drift is the failure mode that matters for series-based content.

Measure prompt fidelity for compound scenes. When the prompt says "two characters, one anxious, one calm, in a rainstorm with distant thunder", how many of those elements land in the output? The advertised strength is compound scene generation; measure it directly.

Benchmark cost against your actual output length. Per-second billing changes the math for shorter outputs. A 15-second ad on Seed Audio at PiAPI's $0.002/sec is $0.03. The same on per-character TTS pricing may be cheaper or more expensive depending on script length.

Watch for latency. All the provider docs describe Seed Audio 1.0 as asynchronous. That is a workflow constraint if you need low-latency, real-time generation.

What Builders Should Do Today

Three concrete actions:

1. Test on your actual scripts, not the demo prompts. The vendor examples ("a suspense radio drama in a late-night convenience store") are the easy case. Compound scene generation with named characters, emotional beats, and specific ambience is where the model has to earn the "audio scene" framing.

2. Pin a provider by price and latency, not by hype. The four listed providers vary by roughly 60% on price. For workloads at scale that is meaningful. Also, given Seed Audio's async model, provider polling behavior matters as much as raw price.

3. Don't trust the "video-conditioned" framing until a primary source confirms it. Third-party blogs describe Seed Audio 1.0 as generating audio from video input. The primary API docs on PiAPI and Atlas Cloud describe text and image/audio reference inputs, not video. If your workflow depends on video-conditioned generation, verify against a provider's actual schema before committing.

The Week Ahead

Three observation signals to track over the next two weeks:

• Watch for an official ByteDance model card or technical report. The Lark wiki is a product doc, not a technical one. An architecture paper would resolve the diffusion-vs-autoregressive question that third parties are already speculating on.
• Run your own five-script eval before relying on any "consistency" or "quality" claim. The community has not yet published a measured comparison; you are unlikely to find a shortcut here.
• Check whether Seed Audio 1.0 pairs with Seedance 2.0 in any lip-sync or sync-conditioned pipeline. ByteDance shipping a matched audio-and-video pair from the same Seed family would be a distinct capability. Nothing in the current signal set confirms that pairing exists yet.

Building similar text-to-audio or dialogue workflows? On kie.ai you can try ElevenLabs V3, Elevenlabs Text to Speech, and OmniHuman 1.5.