SeedRealtime Audio-Visual Full-Duplex LLM Released: Toward Omni-Modal Natural Interaction

ByteDance released SeedRealtime, a native audio-visual full-duplex LLM, and has fully rolled it out for large-scale deployment.

• End-to-end human evaluation shows it cuts audio-visual conversational pacing issues by half versus cascaded models.
• It proactively speaks up when it spots visual changes, like a key target appearing, and can embed tool calls in responses.
• It uses visual context to resolve homophone ambiguity and interpret temporal references in what it sees.
• It distinguishes bystander chatter and background noise to avoid false triggers in noisy environments.

Developers building real-time AI assistants can now deploy a single end-to-end model that watches, listens, and speaks, replacing cascaded ASR-VLM-TTS pipelines and their latency.

Sources

• ByteDance Seed: SeedRealtime Audio-Visual Full-Duplex LLM Released: Toward Omni-Modal Natural Interaction

Read this as text

Back to the AI news