ElevenLabs releases SFX v3 sound effects model: multi-part prompts, better music and vocals, now in ElevenAPI (limited access)

Background and pain point

Games, short videos and ads usually layer several sounds in one scene: a cheering crowd, a ball drop, a commentator shouting. Earlier text-to-sound models handled single sounds best, often followed only part of a complex prompt, distorted vocals or music, and clipped on high-frequency chimes and whistles. ElevenLabs announced SFX v3 on X on Oct 29, saying big scenes rarely come down to one sound, and pitched it as multi-layer sound from a single prompt.

Architecture and mechanism

ElevenLabs has not disclosed the architecture or parameter count. The docs confirm the model ID eleven_text_to_sound_v3 on the same Sound Effects API as v2, so switching only means changing model_id. Compared with v2 it follows several instructions in one prompt (for example "Football crowd cheering, hearing a ball drop, crowd booing"), handles music or vocal prompts such as "Gospel choir singing hello", sounds more natural, and is less likely to clip on gongs, chimes and whistles. It also supports loop for seamless loops.

Benchmarks

No quantitative benchmarks or listening-test scores were published, only demo videos. The docs say results vary by prompt and recommend comparing both models on your own prompts. Positive user reactions on X are personal opinions, not evaluations.

Getting started

Pass model_id: "eleven_text_to_sound_v3" to the Sound Effects endpoint; v2 stays the default, so nothing upgrades automatically. v3 is in limited access: accounts without access get an Unknown model error and some regions are excluded, so keep a v2 fallback. Duration can be 0.1 to 30 seconds, billed at 40 credits per second when specified (endpoint-wide rule). Non-developers can use it in ElevenCreative Sound Effects, Flows and Studio. Describe the action, setting and atmosphere in prompts.