2026-10-27 · VideoFreeTier

Seedance 2.5 Audio Reference: 10 Audio Files, and the Voice Kling 4.0 Only Takes as an ID

Seedance 2.5 takes up to 10 reference audio files, 30 seconds combined, and generates a soundtrack by default. Kling 4.0 takes a voice, not a file: at most 3 voice_id values per task, on one of its three endpoints, from a voice library you have to create first. The same Kling 4.0, published on Replicate, exposes no audio field of any kind.

"Supports audio" turns out to be three separate questions that get collapsed into one adjective. Can the model take audio in? Does it put audio out? Can you turn the audio off? Seedance 2.5 answers yes, yes, yes. Kling 4.0 answers yes to one of them, and which one depends on the surface you read.

This page was first built on 27 October 2026 from platform schemas, because Kling's own API reference rendered entirely in JavaScript and returned an empty page. It renders now. The second reading, on 3 November, replaces the Kling half of the earlier conclusion: not "no audio field anywhere", but "no audio field on the hosting platform, and a voice field, never a file, on the vendor's own API".

Everything below comes from three kinds of page, read without a login and without spending a credit: the vendors' own product pages and the input schemas embedded in model pages on Replicate, both on 27 October, then the vendors' API references on 3 November — Kling's read through a headless browser, fal's decoded from the OpenAPI document embedded in its model page. No clip was generated for this page. That limit is stated again at the end, where it matters.

The three audio questions, separated

A model that writes a soundtrack and a model that listens to one are doing opposite things with the same word. The comparison videos that rank these models measure the first thing — whether the sound is convincing — and stop there. The second thing, whether you can hand the model a voice and have it follow that voice, is a different capability with different failure modes, and it is the one that decides whether a workflow is possible at all.

There is a third question underneath both: whether the audio can be switched off. A model that always renders sound is not the same product as one that renders it on request, and the difference only shows up when you are cutting a clip into a track you already have.

What Seedance 2.5 actually accepts

BytePlus describes the model with a number and a list: "30s narratives, 50 multimodal inputs, and precision editing in one model," and separately, "Bring together images, video, and audio references so the model can follow complex character setups, product details, scene structures, camera moves, and brand assets." The same page promises that it "seamlessly generates synchronized audio and video with millisecond precision across all sonic layers, from ambient effects to human speech."

That is a vendor claim about capability. The field list is where it becomes something you can act on. Seedance 2.5's schema on Replicate defines three reference arrays, each with its own ceiling:

Two constraints sit inside those descriptions and are easy to miss until a request fails. On this platform, audio references cannot stand alone — the schema requires "at least one reference image or video" alongside them. BytePlus's own model documentation says the opposite for its API: Seedance 2.5 "newly supports generating videos with pure audio references, without requiring image or video assets." Both statements are dated and attributable, they disagree, and this page records the disagreement rather than resolving it. The reference arrays are also mutually exclusive with first-frame and last-frame input on Replicate: a clip that starts from image or last_frame_image cannot also carry references. So the 50-input budget is not a bigger version of image-to-video; it is a different mode.

Audio output is a separate switch. generate_audio is a boolean, default true, described as generating "synchronized audio with the video, including dialogue (use double quotes in prompt), sound effects, and background music." Default-on is worth noting for anyone who assumed silence was the resting state.

The three ceilings add up to 30 + 10 + 10 = 50, and BytePlus publishes the same split in its own model documentation: "50 (30 images + 10 video clips + 10 audio clips)". When this page was first written, that arithmetic was labelled this site's, because the product page gave only the total. The API documentation gives the division, so the match is no longer a coincidence — the platform's three ceilings land on the vendor's published allocation.

What Kling 4.0 lists, and on which surface

Kling's own product page is still vague. Its Kling 4.0 block is headed "Audio-Visual Performance for Realistic Results" and reads, in full: "Stable Dynamic Motions and an immersive audio-visual experience, taking video generation to professional-grade production." The word audio appears. What audio means there is never specified — not whether it is generated, not whether it can be supplied, not whether it can be turned off.

The API reference is not vague, and it splits the answer three ways. Kling 4.0 exposes three task endpoints — POST /text-to-video/kling-4.0, POST /image-to-video/kling-4.0 and POST /omni-video/kling-4.0 — published twice, at api-beijing.klingai.com and api-singapore.klingai.com, with the same paths and the same fields. Every one of them takes its inputs as a typed array, contents[]. On the omni endpoint that array accepts nine entry types:

The voice entry is the whole of Kling 4.0's audio input, and its limits are published rather than inferred: "At most 3 voices can be referenced," each one counts toward the combined cap — "The combined number of images, videos, Elements and voices must not exceed 15" — and the value is not a file. "The voice_id is returned by the Custom Voice API, or you can use a system preset voice." So the request carries an identifier pointing at a voice that already exists in Kling's system. Handing over a clip means creating a voice from that clip first, as a separate step, and then referencing the ID.

The other two endpoints carry less. On /image-to-video/kling-4.0 the contents list is a prompt plus first, middle and last frames — no audio entry, no voice entry. On /text-to-video/kling-4.0 it is a prompt. And settings on all three takes only resolution, aspect ratio and duration. There is no audio on/off switch anywhere in the Kling 4.0 request body. Audio is not a field you set; it is something the prompt is expected to describe, which the vendor's prompt guidance calls "audio requirements".

On Replicate the same model is thinner still. Kling 4.0 there defines these inputs:

There is no audio row. Not an audio input, not a generate_audio flag, not a mute parameter. The string "audio" appears exactly once on that page and it belongs to a platform feature toggle in a permissions object, not to the model.

fal, the second hosting platform checked, does not list Kling 4.0 at all: its Kling catalogue tops out at v3 pro and o3 on 3 November. Its Kling V3 (pro) image-to-video endpoint does carry an audio switch — generate_audio, boolean, default true, "Whether to generate native audio for the video. Supports Chinese and English voice output. Other languages are automatically translated to English." No audio input, though. fal's audio input for Kling sits on a separate lipsync endpoint, kling-video/lipsync/audio-to-video, which requires a video_url and an audio_url of 2 to 60 seconds and 5MB at most, and returns a re-dubbed clip rather than a generated one.

The comparison that matters is not Kling against Seedance. It is Kling against Kling. On Replicate the same account publishes Kling 2.5 Turbo Pro and Kling 2.1, and neither of those schemas has an audio field either — three versions in a row expose nothing on that platform. On the vendor's own API the field exists, but as a voice reference on one endpoint out of three. The platform's silence is therefore an integration gap, not a property of the model, and the earlier reading on this page — "Kling's own API documentation would settle it" — has now been settled in the direction that weakens the platform conclusion rather than strengthening it.

The complaints that now have a mechanism

Two comments under the comparison videos describe symptoms that the field lists explain, at least partly.

"Kling 4 doesn't support audio reference (for character voices) on any platform I used it on, which is a bummer." — @@vfxfan, YouTube comment, 2 likes, read 27 October 2026

"I also find Kling 4.0 very plastic. But the worst was that it adds music even if you write no music. So for cinematic workflow is not good." — @@PyrosAICreative, YouTube comment, 1 like, same date

The first is a user's survey of the platforms they happened to try, and the vendor's own reference now explains the shape of it: none of the surfaces read here accepts an audio clip as a reference. Kling 4.0 accepts a voice ID, and getting your own voice into that library is a separate step through the Custom Voice API. "No audio reference on any platform I used" and "voice reference on the omni endpoint" can both be true, because they describe different objects.

The second is a complaint about output, and the request body shows a plausible route: with no audio switch in settings and no generate_audio field on the platform, nothing in the request can carry a "no music" instruction except the prompt itself. That is a mechanism worth testing, not a proven cause — one generation with no music, silent in the prompt would show whether the audio track survives it, and that generation was not run here.

What this does not settle

No clip was generated for this page. Whether Seedance 2.5's soundtrack is any good, whether Kling 4.0 adds music when told not to, and whether either model's audio track is present at all in a delivered file are all unmeasured here. Reading a schema tells you what a request can contain; it does not tell you what comes back. Anyone quoting this page for audio quality is quoting something it does not say.

Most of the field readings are platform-side. Replicate publishes both models under the vendors' own accounts, which is why the two sides are comparable at all, but a field list on a hosting platform is an integration, not a specification — and the Kling 4.0 reference shows how much an integration can drop. What the vendor's reference still does not say is whether a referenced voice holds: whether timbre, accent and pacing survive a 30-second clip is unmeasured here, because no clip was generated. Seedance 2.5's 10-audio ceiling remains this channel's ceiling; the vendor's own reference documents the audio role separately, and that role is not the same object as Kling's voice ID.

The one thing the two sources disagree on is whether an audio reference needs an image or a video beside it. This page records both readings and does not pick a winner; one request would settle it, and none was sent.

And the reference limits are ceilings, not recommendations. Seedance 2.5's thirty images plus ten videos plus ten audios is a legal request, not a tested one, and so is Kling 4.0's fifteen-item cap with three voices inside it — no request at either size was sent, so nothing here describes what happens at the boundary.

Prices, tiers and model versions on both sides move without announcement. The readings behind this page are dated 27 October and 3 November 2026 and carry the channel each was taken from; re-read before quoting.

Related on this site: Kling 4.0 vs Seedance 2.5: what five seconds actually costs — the money side of the same pair, with the tier arithmetic. Seedance 2.5 daily generation cap — how many clips the free allowance actually produces. Video to video AI free — video as a motion reference, which is the neighbouring input type.