ByteDance's Seed lab has released Seed Audio 1.0, an audio generation model that can produce dialogue, sound effects, and ambient audio from a single prompt — and control when each sound enters along a timeline.
The model accepts text and reference audio inputs, with time control precision down to 100 milliseconds. It can generate roughly two minutes of audio in one go, and can extend recordings while keeping character voices consistent.
Seed Audio 1.0 supports over 20 languages. The same voice can be transferred across languages, and its tone and emotion can be adjusted.
In official evaluations, the audio was deemed usable in over 90% of tested scenarios, and most languages scored above 4 on the Mean Opinion Score (MOS) for naturalness. The model is now available on the Volcengine experience center, with API access open.