ByteDance Releases Seed Audio 1.0 with Fine-Grained Temporal Control for Cinema-Grade Sound Creation

ByteDance released Seed Audio 1.0 on July 20, marking a new stage in AI audio generation technology that moves beyond single speech synthesis to complete sound scene creation. The model is now available for testing at the Volcano Fangzhou Experience Center.

Traditional cinema-grade audio production requires separate models for vocals, sound effects, and background music, followed by manual editing and mixing — a lengthy process that struggles to maintain narrative coherence. Seed Audio 1.0's core breakthrough is not simply concatenating materials but jointly modeling various audio elements within a unified framework, generating complete sound works that serve storytelling end-to-end.

The model offers three core capabilities: precise temporal-spatial arrangement, supporting 100-millisecond accuracy for dialogue and sound effect timing along the timeline, ideal for video dubbing and ad production; stable voice performance with zero-shot generation and long audio extension, maintaining character voice consistency while naturally expressing different emotions including anger and joy, even allowing the same voice to perform multiple characters; and fluent multilingual support covering more than 20 languages including Chinese, English, and Japanese.

Evaluation data shows the model's audio availability exceeds 90% across nine common creative scenarios, with MOS scores for multilingual generation consistently above 4.0 (excellent level).

On the BytePlus platform, Seed Audio 1.0 supports multiple input modes including text prompts (up to 3,000 characters), reference audio, and reference images, generating up to 120 seconds of audio per segment with extension capabilities for longer formats. BytePlus positions it as the first reference-based audio generation model in its Dola product line.

From Seed-TTS to Seed Audio 1.0, ByteDance's audio AI roadmap is progressively unfolding. The team plans to further integrate multimodal inputs such as video references and explore controllable translation technology, continuously optimizing long-form audio and track generation capabilities.