SeedAudio1.0 generates dialogue, effects, and ambience in one shot

智东西 · wechat · 2026-07-20

ByteDance has released SeedAudio1.0, an audio creation model that can generate a full sound scene — including dialogue, sound effects, and ambient audio — from a single prompt.

The article’s hands-on tests suggest the model can already do several things well:

generate complete narrative scenes from environment descriptions alone
keep character voices consistent across a long clip
mimic reference voices, including a news-anchor-style female voice
handle multilingual prompts, including Cantonese
combine voice lines with environmental sounds and action cues in a time-aligned way

According to the official evaluation numbers quoted in the post:

it improves subjective preference scores versus top commercial audio services by up to 0.79 CMOS
it reaches 91% usability across more than ten scenarios such as film, short drama, podcasts, and animation
it supports 20+ languages, with most languages scoring above MOS 4.0 for naturalness

The article also highlights two core capabilities:

Multi-element orchestration — the model can place voices, background atmosphere, and effects on a timeline without stitching multiple tools together
Fine timing control — it supports roughly 100 ms-level precision, which makes it more suitable for video dubbing and ad production

The author concludes that AI audio is moving from “can speak” to “can create.”

Original post →

SeedAudio1.0 generates dialogue, effects, and ambience in one shot

More from Multimodal