ByteDance Seed unveils scene-level audio model
ByteDance Seed's Seed Audio 1.0 generates speech, sound effects and ambience together from a single prompt, so creators can produce a whole scene at once instead of layering separate clips on a timeline.
ByteDance's Seed team introduced Seed Audio 1.0 in a blog post dated July 20, 2026. Instead of producing one isolated asset at a time, the model treats the scene as its unit of output. It generates dialogue, sound effects, ambience and background texture together, using a shared context for timing, tone and acoustic space.
A prompt can specify the speaker, emotional tone, line delivery, environment, key sound effects and how the scene unfolds. Prompts can also set when each line of dialogue enters, with timing precision at 100 ms intervals. ByteDance pitches this feature for video dubbing, re-voicing and advertising.
Voices can be defined from a text description, an authorized reference sample, or both, without training a separate model for each speaker. The model can generate up to two minutes of audio in a single pass and can extend it further. ByteDance says it keeps a character's voice consistent across those extensions and across more than 20 languages, including Chinese, English, Japanese, Korean, Spanish, Indonesian, German, French, Thai and Vietnamese.
The architecture pairs a language model with a diffusion-based acoustic generator. The language model turns creative intent into scene-level controls, and the generator produces audio in a high-fidelity latent space that responds to instructions about character, emotion, timing and scene context.
The available material does not include availability, pricing, API access, model size or public benchmark comparisons.