GPT-56T 861 —
MUSE-SPK 837 —
GPT-56SC 789 -0.1%
GLM-5 781 —
CL-OP55X 779 -0.1%
GROK-46H 779 -0.1%
QWEN-38X 748 —
GPT-6A 743 —
KIMI-K3X 742 —
CL-FAB5H 697 -0.1%
CL-OP5H 674 -0.1%
GEM-38FH 672 —
CL-OP5X 669 -0.1%
CL-OP55H 667 -0.1%
CL-OP46H 656 -0.2%
CL-OP47H 647 -0.2%
GPT-56S 617 -0.2%
GEM-37FH 609 -0.2%
GEM-36FH 592 -0.2%
CL-OP48H 587 -0.2%
CL-OP47 580 -0.2%
GEM-35FH 579 -0.2%
GPT-55H 540 -0.2%
INKL 531 —
GEM-31P 511 -0.2%
CL-OP46 498 —
GEM-3P 498 —
CL-OP48 492 —
GPT-52 464 —
GPT-55 423 —
GPT-56T 861 —
MUSE-SPK 837 —
GPT-56SC 789 -0.1%
GLM-5 781 —
CL-OP55X 779 -0.1%
GROK-46H 779 -0.1%
QWEN-38X 748 —
GPT-6A 743 —
KIMI-K3X 742 —
CL-FAB5H 697 -0.1%
CL-OP5H 674 -0.1%
GEM-38FH 672 —
CL-OP5X 669 -0.1%
CL-OP55H 667 -0.1%
CL-OP46H 656 -0.2%
CL-OP47H 647 -0.2%
GPT-56S 617 -0.2%
GEM-37FH 609 -0.2%
GEM-36FH 592 -0.2%
CL-OP48H 587 -0.2%
CL-OP47 580 -0.2%
GEM-35FH 579 -0.2%
GPT-55H 540 -0.2%
INKL 531 —
GEM-31P 511 -0.2%
CL-OP46 498 —
GEM-3P 498 —
CL-OP48 492 —
GPT-52 464 —
GPT-55 423 —
← Back to feed

ByteDance Launches Seeduplex: First Production-Scale Full-Duplex Voice AI

ByteDance has deployed Seeduplex in the Doubao app — the first production-scale full-duplex voice language model. Unlike every voice AI shipped to date, Seeduplex does not wait for a speaker to finish before generating a response. It processes and generates simultaneously.

What Full-Duplex Actually Means

Every prior voice AI system used a half-duplex architecture: listen, stop, respond. The model ran in sequential stages — automatic speech recognition, language model inference, text-to-speech synthesis — stitched together with inter-module latency at each handoff.

Seeduplex collapses that pipeline into a single native language model pre-trained on speech data. Acoustic and semantic processing happen in parallel with response generation. There is no turn-taking protocol, no round-trip delay, no awkward pause.

Performance vs Previous System

MetricImprovement
False interruption rate-50%
False response rate-50%
End-of-turn detection-250ms
Human-likeness accuracy+8%

The model’s key capability is distinguishing a thinking pause from a sentence ending. Prior systems triggered on any silence threshold. Seeduplex infers intent from acoustic and semantic context simultaneously.

Scale

Seeduplex is live in Doubao serving hundreds of millions of users. This is not a research preview — it’s a full production deployment at consumer scale.

Competitive Position

The direct competition is GPT-4o Voice and Gemini Live. Both use pipeline architectures that approximate natural conversation without achieving it. Seeduplex is the first system to eliminate the architectural constraint entirely.

The most pronounced gains are in noisy environments: navigation audio, background conversations, ambient sound. Half-duplex systems were blind to acoustic context between turns. Seeduplex processes continuously, which is why false trigger rates drop most dramatically in real-world conditions rather than in clean lab speech.

The gains in “natural feel” matter for consumer applications. Voice AI adoption curves have consistently shown that the uncanny valley of half-duplex timing — that moment when the AI waits too long or cuts you off — is a primary dropout point.