ElevenLabs SFX v3: Multi-Sound Prompts, Cleaner Audio

A football clip rarely needs one sound. The crowd cheers, a ball drops, then the same crowd boos. With older text-to-SFX models, packing that into one prompt often meant generating pieces and stitching them by hand. On October 29, ElevenLabs answered with SFX v3.

The company blog keeps the claims tight: the new model follows multi-instruction prompts more reliably, handles music and vocals more convincingly, and sounds cleaner than SFX v2 in ElevenLabs' benchmarks, with fewer harsh artifacts on gongs, chimes and whistles. It is live in ElevenCreative — Flows, the Sound Effects tab, Studio and the Music Editor — plus the mobile app. Developers reach it through ElevenAPI with model_id eleven_text_to_sound_v3.

One prompt for a whole scene's audio

SFX v3 is not another preset library. It is a bet that creators can describe a scene's sounds in natural language and hear them land together. Docs push action, setting, atmosphere and performance over category words — “rain against a window in a quiet apartment” beats “rain”; “a couple arguing in French” beat “fighting” in their own examples. Exact counts such as “triangle hit 4 times” are still unreliable.

Clips top out around 30 seconds, with MP3 for all effects and 48 kHz WAV for non-looping ones. Longer atmospheres can use seamless loops. Public API pricing lists sound effects at about $0.12 per minute, with per-second credit charges when duration is set. Those are list figures; real invoices still follow the account plan.

Cleaner on their scorecard — still their scorecard

ElevenLabs says SFX v3 beats v2 on internal benchmarks for multi-sound following, music/vocals and high-frequency cleanliness. The full benchmark tables are not published, and this article did not regenerate side-by-side samples. Creators should run their own house prompts through both models before switching. Docs also flag limited access: some regions or accounts get an Unknown model error.

Video and game pipelines are pulling the SFX category forward. A model that can sketch a whole scene from one prompt saves the fiddliest hours on an edit timeline; it does not replace dialogue design, music composition or a final mix. ElevenLabs is shipping Creative UI and API on the same cadence — hear it in the product, then drop the same model into a pipeline.

SFX v3 nudges text-to-SFX from “one sound at a time” toward “hear the scene together.” Whether that nudge is worth the switch will not be decided by launch copy. It will be decided by the next project's prompts, and by which model makes you redo fewer takes.