ElevenLabs announces the release of SFX v3, its latest sound effects generation model, available in its Sound Effects API. According to the company, it follows multiple prompts better than SFX v2 and handles voices and music more effectively.
Switching to the new model
To use SFX v3, simply change the model identifier, the model_id parameter, setting it to eleven_text_to_sound_v3 on the POST /v1/sound-generation endpoint. An endpoint is the address to which an application sends its requests in an API, i.e., a programming interface.
The duration does not need to be specified: the model chooses the length of the effect itself based on the prompt. If precise timing is required, the duration_seconds parameter can be used to set the desired duration.
When the result deviates from the request, the prompt_influence parameter can be increased beyond its default value, set at 0.3. Higher values make the prompt be followed more closely and reduce variations between generations. The output_format query parameter defines the codec and sampling rate of the file, with formats such as mp3_44100_128 or pcm_48000.
Finally, the looping function currently only works with eleven_text_to_sound_v2. A request requiring a loop must therefore remain on this model, with the loop: true parameter.
Formulating prompts
SFX v3 is presented as particularly effective for multi-sequence sound effects. The order of the prompts matters: events must be listed in the order they should occur, for example “Footsteps on gravel, then a metallic door opens”. For a single sound, a short and concrete prompt yields the best results.
Wording also matters. According to ElevenLabs, one of its testers found that “a couple arguing in French” gave much better results than “a couple fighting in French”. The company recommends testing several phrasings for the same event before finalizing the choice.
Comparing v2 and v3
SFX v3 is described as an incremental update. It is said to be less likely to clip high-pitched sounds, such as gongs, chimes, or kettle whistles, i.e., distort them when the signal exceeds the maximum allowable level. The company’s engineering team also considers the overall rendering to be more natural.
Since both models operate on the same endpoint, it is easy to compare their results: simply send your most frequent requests to both v2 and v3 before switching models. Sound effects are also accessible without code, in the Sound Effects tab, as well as in Flows and Studio, within ElevenCreative.
