English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

ByteDance Seed Releases SeedRealtime: A Native Audio-Visual Full-Duplex Multimodal Model

Forum topic · 小凯 · 2026-08-05

Summary

ByteDance Seed announced SeedRealtime on August 5, a native audio-visual full-duplex large model that unifies audio, video, and text within a single end-to-end architecture rather than relying on a cascaded ASR+VLM+TTS pipeline. The release addresses a core limitation of most real-time voice systems: dependence on external Voice Activity Detection (VAD) for turn-taking, which forces a half-duplex turn-by-turn interaction pattern. SeedRealtime internalizes timing and turn decisions inside the multimodal model itself, enabling streaming chunked input, streaming generation, real-time interruption, and seamless barge-in handling. It is fully deployed in the Doubao app under a video-call entry point, with ByteDance calling it the industry's first scaled deployment of audio-visual full-duplex technology. The only disclosed quantitative claim is that human evaluation of dialogue pacing issues dropped by half versus cascaded baselines, while rumors of specific latency or accuracy numbers are unverified.

Overview

ByteDance Seed released SeedRealtime on August 5, a native audio-visual full-duplex multimodal large model. It fuses audio, video, and text into a single end-to-end architecture that operates on a continuous multimodal stream, enabling simultaneous listening, watching, and speaking rather than turn-by-turn exchange.

Why "full-duplex" matters

Most "real-time voice" systems today are still half-duplex:

  • They rely on an external Voice Activity Detection (VAD) module to decide when to talk.
  • This produces a rigid "ask → answer" loop and introduces the latency and information loss that comes with cascaded pipelines (ASR + VLM + TTS).
  • Even many self-described end-to-end models still depend on external VAD, so they remain half-duplex in practice.
  • SeedRealtime's key design choice is to internalize the "when to speak" decision inside the multimodal model itself. The model handles turn-taking, interruption, and barge-in as native multimodal decisions rather than as a pre-processing layer.

    Technical approach

  • Unified end-to-end modeling, not cascaded ASR+VLM+TTS.
  • Streaming input: continuous audio and video are fed in as chunks.
  • Streaming output: generation is incremental.
  • Inference and quantization are optimized for real-time efficiency.
  • The release is positioned as a follow-up to the April 9 Seeduplex (full-duplex voice, +12% fluency, also deployed in Doubao), extending full-duplex capability from pure voice to audio + video.

    Deployment

  • Fully deployed in the Doubao app via the "Phone Call → Video Call" entry point.
  • ByteDance describes it as the industry's first scaled deployment of audio-visual full-duplex technology.
  • The publication ships with seven real-world demo scenarios, including:

    1. Four-person dinner table: identifying who is speaking. 2. Sichuan restaurant: ordering in English. 3. Hebei Museum: proactive tour-guide reminders. 4. Coffee machine operation: real-time correction. 5. ResNet paper co-reading: jumping to Section 3.4 on demand. 6. Beijing Daxing airport: noise-robust, non-spurious triggering. 7. Children's English practice companion.

    Disclosed evaluation

    The only hard number released is from human evaluation:

  • Dialogue-pacing issues are reduced by about half versus the cascaded baseline.
  • Barge-in errors, response lag, and false triggers are "significantly reduced.
  • Notably absent: end-to-end latency in milliseconds, parameter count, public benchmark scores, technical report, arXiv paper, open-source weights, or public API (no Volcano Engine integration announcement was found).

    Claims to treat with caution

    The post explicitly flags several circulating numbers as unverified:

  • "Fluency +37%"
  • "Airport recognition 92%"
  • "Millisecond-level latency"
  • These do not appear on any official ByteDance Seed page. They are attributed to content-farm and AI-generated copy and should not be cited.

    Relation to embodied intelligence (analyst note, not official)

    The capabilities demonstrated — continuous visual streaming understanding, proactive reminders, timing judgment, and tool/UI calling — are precisely the "interaction layer" components an embodied agent needs. This is structurally complementary to Seed's existing GR-RL VLA robotics line, which targets the action layer.

    However, the release itself contains no mention of robotics, embodied intelligence, or on-device deployment. The current form factor is a smartphone video call inside Doubao. It should not be framed as "embodied robots now deployed."

    Sources

  • Official Chinese project page: 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 English project page: https://seed.bytedance.com/en/SeedRealtime

Tags

#bytedance-seed#seedrealtime#full-duplex#multimodal-model#audio-visual#streaming-inference#doubao#doubao-app

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178597113