Voice and audio
ElevenLabs SFX v3: multi-part sound effects from one prompt
ElevenLabs has released SFX v3, a sound effects model that follows prompts with several sounds in sequence more closely, and handles music and vocals better. It is available in the app and through the API.
ElevenLabs has released SFX v3, the newest version of its text-to-sound-effects model. The company published the developer announcement on 2 November, and the model is live in its ElevenCreative app and in the API.
What changed
ElevenLabs says SFX v3 follows prompts that describe several sounds more reliably, so more of what you ask for ends up in the result. Its own example is a football scene: a crowd cheering, a ball dropping, then the same crowd booing, all from one prompt.
It also says the model generates music and vocals more convincingly, and that audio is cleaner than SFX v2 in its benchmark testing. The company points to fewer artefacts on sounds such as gongs, chimes and kettle whistles, which are prone to clipping.
ElevenLabs calls v3 an incremental update, and recommends running your usual prompts through both versions before switching.
How to use it
In the API, set model_id to eleven_text_to_sound_v3 on the existing sound-generation endpoint. Both models share that endpoint, so comparing them is a one-line change. In ElevenCreative, the model is available in the Sound Effects tab, Flows, Studio and the Music Editor, and in the mobile app.
The developer guide offers some practical advice:
- List events in the order they should happen, for example footsteps on gravel, then a metal door opening.
- Keep single sounds short and concrete.
- The model chooses a length from the prompt, or you can set
duration_seconds. - If results drift from the prompt, raise
prompt_influenceabove its default of 0.3.
Limits
Seamless looping currently works only with the v2 model (loop: true), so ambience beds still need v2. The documentation lists a maximum of 30 seconds per generation. We have not tested v3 ourselves, and every quality claim above is ElevenLabs’ own.
Our take
Sound is the cheapest way to make a generated world feel real, and it is usually the last thing teams budget for. A model that handles a sequence of sounds from one sentence suits prototyping: describe a scene and hear a rough pass in seconds. For anyone building interactive worlds, the open question is runtime. A 30-second file is an asset, not a living soundscape, and the games described in our essay on AI models as game generators will need audio that reacts as the world changes.