ByteDance's Seed team released SeedRealtime on August 5 — a natively full-duplex audio-video large language model. The key difference from most "realtime voice" models: instead of relying on an external VAD (voice activity detector) to decide turns in a half-duplex, one-question-one-answer loop, SeedRealtime puts "when to speak" inside the model itself.
What it is
- Unified architecture: audio, video, and text are fused into a single end-to-end model, interacting in real time over a continuous multimodal stream — "watching, listening, and speaking" simultaneously.
- Not a cascade: the official announcement explicitly calls out the latency accumulation and information loss of ASR+VLM+TTS pipelines, and notes that most end-to-end models still depend on external VAD and are therefore essentially half-duplex.
- Native turn-taking: conversational turn decisions are internalized as the model's own multimodal decisions, enabling real-time interruption and natural hand-off.
- Streaming design: continuous audio-video input in chunks, streaming generation output, with efficiency from quantization and inference optimizations.
- Continuous visual understanding + proactive reminders + timing judgment + tool calling map naturally onto the "interaction layer" of embodied agents, and could complement Seed's existing GR-RL VLA robotics line — but this article's connection is our inference, not an official claim. The announcement never mentions robots, embodied intelligence, or on-body deployment; the current form factor is a video call inside a phone app.
- There is no technical report, no arXiv paper, no open weights, and no API (no Volcano Engine integration announcement), so third parties cannot reproduce or independently evaluate the system.
- The "reduced by half" figure comes from the vendor's own human evaluation; evaluation-set size and baseline configuration are undisclosed.
- Rumored figures circulating online — "37% fluency improvement," "92% airport recognition," "millisecond-level latency" — have no source on the official pages. Content-farm articles and apparent AI-generated writeups are unreliable; do not cite these numbers.
- Official project page (Chinese): https://seed.bytedance.com/zh/SeedRealtime
- Official English blog: https://seed.bytedance.com/en/blog/seedrealtime-audio-visual-full-duplex-llm-released-toward-omni-modal-natural-interaction
- Official project page (English): https://seed.bytedance.com/en/SeedRealtime
Deployment
SeedRealtime is fully rolled out in the Doubao app (chat box → "call" → video call entry). ByteDance claims it is "the first at-scale deployment of audio-video full-duplex technology in the industry." The predecessor is Seeduplex (released April 9, full-duplex voice-only, +12% fluency, also on Doubao); SeedRealtime extends that line from voice to audio-video.
Metrics
The only official quantitative claim: in end-to-end human evaluation, conversation-rhythm problems were reduced by half versus cascade models, with significantly fewer interruptions, hesitations, and false triggers. No latency in milliseconds, no parameter counts, no public benchmark scores.
Seven real-scenario demos accompany the release: identifying people and voices at a four-person dinner, ordering in English at a Sichuan restaurant, proactive reminders during a museum tour in Hebei, real-time correction while operating a coffee machine, guided reading of a ResNet paper (locating Section 3.4), noise-robust use at Daxing Airport without false triggers, and children's English tutoring.