Kuaishou releases Kling 2.6 with native audio

Background

AI video generators have mostly been asked to do one thing at a time: turn a text prompt into moving pictures. Voice, singing and background sound were handled elsewhere — by a separate speech tool, by a sound designer, or by someone in an editing suite — and then laid over the finished footage. The split shows in the workflow: a few seconds of generated footage can pass through several tools before it is watchable with sound.

What was released

On 3 December 2025, Kling AI released Video 2.6 ^[1]. The central feature is what the company calls simultaneous audio-visual generation: the picture and the sound are produced in the same generation rather than in separate steps ^[1]. The audio is not limited to music or background noise. According to the launch material, the model generates speech, singing and ambient effects ^[1]. Voices are offered in Chinese and English ^[1], and a single clip runs up to 10 seconds ^[1].

Where it sits in the launch

The model was released as part of Omni Launch Week ^[2]. Kuaishou's launch release positioned it for advertising, social media and e-commerce content ^[1] — three formats where short, frequent, sound-on video is the everyday unit of work. The company also said the model holds a "world-leading position in Chinese voice generation performance" ^[1]. That phrasing is the developer's own characterization of its model, offered in the announcement rather than as the result of an outside comparison.

Why generating both at once matters

The usual reason to separate picture and sound is that they are different problems: one is about motion, framing and consistency across frames, the other about timbre, prosody and timing. Keeping them apart is tidy for engineers but expensive for users, because every handoff adds a step where the two tracks can drift out of step — a line of dialogue that no longer fits the gesture it was written for, a sound effect landing a beat late. When speech and ambience come out of the same pass as the images, the timing of the two is decided together instead of being repaired afterwards ^[1]. For the formats Kuaishou names — ads, social posts, product video — that matters, because those are short pieces that live or die on whether a spoken line and the shot it belongs to feel like one thing.

Reading the details

The 10-second clip length is worth noticing ^[1]. It sets the unit of work at a shot rather than a scene: a spoken sentence, a snatch of song, a street corner with its own noise. Nothing in the announcement suggests the model is meant to carry a long narrative on its own; the practical picture is a tool that fills in the sound-bearing fragments that make up a short video, which is consistent with the advertising, social and e-commerce uses Kuaishou pointed to ^[1].

The Chinese-and-English voice support also says something about the intended users ^[1]. A model that can speak both languages is aimed at creators working across markets, or at teams producing one campaign in two versions.

Open questions

The launch material is a product announcement, so it answers the what and not the how. It does not report how the model was trained, what data the voices were built from, or how it compares with other systems on a shared test — the claim about Chinese voice performance is stated without a benchmark attached ^[1]. Those are the questions that will decide how the model holds up in daily use, and they will be answered by users rather than by the release itself.