ByteDance ships full-duplex voice model to Doubao
ByteDance Seed has fully rolled out Seeduplex, a native full-duplex speech LLM, on the Doubao app, bringing listen-while-speaking voice AI out of the lab and to hundreds of millions of users.
ByteDance's Seed team introduced Seeduplex, a speech LLM that listens and speaks at the same time. It replaces Doubao's previous end-to-end speech model, which worked in a half-duplex way, with users and the model taking turns.
Seed says older half-duplex systems usually chained separate modules together, such as standalone voice activity detection (VAD) to split audio and traditional noise reduction. Because each module saw only acoustic cues or local text, these systems were easily thrown off by noisy rooms or answered while a user was still pausing. Seeduplex is built on a proprietary LLM foundation, pre-trained on speech data, and models sound and meaning together.
Seed reports these gains over its half-duplex model:
- Interference suppression: false responses and false interruptions roughly halved in complex acoustic scenarios.
- Endpoint detection: premature responses down 40%. In a human-likeness test focused on endpoint detection, the model scored 8% better.
The company gives in-car navigation prompts, side conversations in a café and talking to a delivery person as cases where the model should ignore speech not aimed at it. Seed also says it worked through engineering problems such as latency spikes and stability under heavy concurrent load.
Seed says Seeduplex "notably outperforms" voice calling in mainstream apps, but the material provided does not name those apps, share benchmark details or describe the model's size or architecture. A project page is available. The post does not say whether the model will be offered outside Doubao.