Seed Audio 1.0

Overview

Seed Audio 1.0 is ByteDance's all-in-one audio model, built as the next step for text-to-speech rather than a departure from it. Traditional audio workflows use separate tools for voice, music, and sound effects, each generated independently and then mixed. Seed Audio 1.0 produces all three together from a single text prompt, and the output is a complete, ready-to-use audio track.

From text-to-speech to reference-to-audio

Ordinary TTS takes a voice and a block of text. Seed Audio 1.0 takes a scene: the weather and the room, the music underneath, the effects around the action, what each character looks and sounds like, and the lines they speak. Everything you can describe becomes part of the same generation, which is why a prompt can run to 3,000 characters without being padding.

The model understands scene context. A café conversation prompt gets ambient chatter and the right acoustic space, not generic room tone. An action cue gets sharp effects and a dynamic score, not background music pasted over silence. The three layers come out proportioned and mixed for the scene, not assembled from separate exports.

Four modes

The simplest is text-to-speech: pick a voice, enter text, and the model reads it. Voice cloning adds one step, upload a single audio clip and the cloned voice becomes available for that same text-to-speech. The other two modes are where the model separates itself from ordinary TTS.

In text prompt to audio (T2A), every voice comes from your description. Name a character's age, accent, emotion, tone, and speed, and the model casts it.

In text prompt plus audio to audio (TA2A), you upload up to three reference clips of up to 30 seconds each and tag each one to a character in the prompt. The generated voice then follows the recording rather than a description. Reference clips can be uploaded for a single job or kept in an asset library and reused across a series.

Twenty languages

Seed Audio 1.0 generates in English, Chinese, Japanese, Korean, Mexican Spanish, Castilian Spanish, Indonesian, German, Brazilian Portuguese, French, Thai, Vietnamese, Malay, Filipino, Italian, Russian, Dutch, Polish, Turkish, and Swedish. Prompt language and script language should match: write the brief in the language the characters speak.

Timing you can specify to the second

Seed Audio 1.0 accepts a timestamp on any line, written inline as [5.5s:8.0s], and fits the delivery to that exact window. For dubbing, this is the difference between audio that roughly fits and audio that lands on the cut. Generation is non-streaming: the model renders the finished track in a single pass, up to two minutes long.