Runware adds ByteDance's Seed Audio 1 model
By Steven Van ·
The API generates speech, music, and sound effects from one prompt, and can derive a character voice from a portrait image.
ByteDance's Seed Audio 1 has arrived on Runware, giving developers a fast, low-cost API path to one of the most capable audio-generation models released this year. Seed Audio 1 is not a plain text-to-speech engine — from a single text prompt it produces natural speech, sound effects, music, and ambient sound, and mixes them into a finished audio scene. Built by ByteDance's Seed team and launched in late June 2026, the model is now rolling out across major inference platforms, and Runware's addition brings it to teams that want to run it at scale without provisioning their own GPUs.
Dialogue, SFX, Music, and Ambient Sound From a Single Prompt
The headline capability is breadth. Where most audio models specialize in one output type, Seed Audio 1 treats an entire soundscape as a single generation. Describe a scene in text — a conversation over a rainy street with a distant siren and a soft piano bed — and the model returns speech, foley-style effects, music, and environmental ambience together, already mixed. For creators building video, games, or interactive experiences, this collapses a workflow that normally spans a TTS tool, a music generator, and a sound-effects library into one API call. Running that through Runware means the heavy inference happens on managed infrastructure rather than hardware you have to stand up and maintain.
Multi-Speaker Conversations With Distinct Character Voices
Seed Audio 1 generates multi-speaker dialogue in a single pass, assigning each speaker a distinct voice rather than looping the same synthetic voice through different lines. That makes it practical for scripted scenes, podcast-style exchanges, explainer videos with a host and guest, or NPC banter in a game — all rendered coherently so voices stay consistent across a conversation. Because the speakers are handled together in one generation, timing and turn-taking feel natural instead of stitched, which is exactly the kind of output that is tedious to assemble by hand from separate clips.
Voice Cloning From Up to Three Reference Clips
On Runware, Seed Audio 1 supports voice cloning from up to three reference audio clips. Supply short samples of a target voice and the model adapts its output to match that speaker, so you can keep a consistent character or brand voice across a whole project without re-recording. This zero-shot-style adaptation is one of the model's core strengths — it also underpins cross-lingual dubbing, where a cloned voice can carry the same identity into another language. For localization and dubbing pipelines, that turns per-language voice casting into a reference-clip step rather than a full recording session.
Portrait-to-Voice: A Character Voice Derived From an Image
Perhaps the most distinctive feature is portrait-guided voice: Seed Audio 1 can derive a character voice from a single portrait image. Instead of hunting for a reference recording, you give the model a face and it infers a fitting voice for that character. Paired with image and video pipelines — the kind Runware already hosts alongside ByteDance's visual models — this opens a route to fully synthetic characters that look and sound consistent from one asset. It is a natural fit for animation, avatars, and character-driven storytelling where you have a visual identity before you have a voice.
For developers, the appeal is as much about delivery as capability. Runware positions itself as a single, high-throughput inference API spanning image, video, and audio models, with usage-based pricing and no GPUs to provision — the company advertises rates well below typical market cost. Bringing Seed Audio 1 onto that stack means the same account and endpoint that already runs ByteDance's visual models can now generate the audio to match, which is a meaningful simplification for anyone shipping multi-modal features. Seed Audio 1 is landing across the ecosystem — fal, WaveSpeed, Runway, and Higgsfield among them — and Runware's API-at-scale angle makes it a strong pick for production workloads.
Sources: Runware — ByteDance models, AlphaSignal, ByteDance Seed — Speech.