fal adds ByteDance's Seed Audio 1.0 to its inference API
By Steven Van ·
Developers get a metered, text-to-audio endpoint for speech, music, sound effects and soundscapes, no Volcano Engine account needed.
ByteDance's Seed Audio 1.0, the Seed team's all-in-one audio-generation model, is now runnable on fal's fast inference API, giving developers a single endpoint for speech, music, sound effects, and full soundscapes.
Seed Audio 1.0 now runs on fal's inference API
fal announced Seed Audio 1.0 on June 25, 2026 with a simple invitation to "Try it here today!" The model is listed as a partner offering with commercial-use rights, exposed through the same text-to-audio endpoint pattern fal uses for its image and video models. That means the ByteDance model that debuted days earlier at the Volcano Engine FORCE 2026 conference is now available to any developer with a fal API key, no separate ByteDance or Volcano Engine account required. Pricing on fal is metered per minute of generated audio, so teams pay for output rather than idle GPU time.
What Seed Audio 1.0 generates from a single prompt
Unlike a traditional text-to-speech engine, Seed Audio 1.0 produces the full spectrum of audio from text. From one prompt it can generate natural speech, original music, foley-style sound effects, and ambient environmental soundscapes, and it can stack them together in a single pass. Across the broader ecosystem the model also supports multi-speaker dialogue with distinct voices per character, zero-shot voice cloning from short reference clips, character voices guided by a portrait image, and multilingual output and dubbing, with the wider Seed Audio rollout targeting up to 18 languages. On fal's endpoint today the model accepts text, reference audio, and an image as inputs, generating up to two minutes of audio per pass with English and Chinese available now and broader language support slated for the July 2026 update.
Building audio into apps with fal's speed and scale
fal's whole pitch is fast, scalable inference for generative media, and putting Seed Audio 1.0 behind that infrastructure is what makes it interesting for builders. Instead of standing up their own audio stack, developers can call Seed Audio 1.0 the same way they already call image and video models on fal, batching requests and scaling on demand. The workflow guide fal published covers ten patterns, including sound-design-heavy scenes, text-to-speech with saved voice libraries, extending existing clips, inpainting to fill gaps, stitching multiple takes together, and re-editing dialogue. Reference voices are tagged inline as @Audio1, @Audio2, and @Audio3, with up to three 30-second clips and prompts up to 2,048 characters, so a single API call can direct a full multi-character scene.
Where Seed Audio 1.0 on fal fits the AI-audio landscape
fal is one of several inference platforms picking up Seed Audio 1.0 as ByteDance rolls it out, alongside WaveSpeed, Runware, Runway, and Higgsfield. Runway, for instance, is positioning the model for cinematic soundscapes inside video workflows. fal's differentiator is its developer-first, speed-focused API rather than an end-user creative suite, which suits teams embedding generated voiceover, dynamic game or app audio, podcast and dubbing pipelines, or scored video into their own products. For creators and studios, that combination of a genuinely multi-modal audio model and fal's low-latency, pay-per-output infrastructure lowers the barrier to shipping AI audio features without managing model hosting themselves.
Sources: fal.ai/models/bytedance/seed-audio-1.0, fal.ai/learn/tools/how-to-use-seed-audio, alphasignal.ai.