ElevenLabs releases SFX v3: the model follows multi-sound prompts more accurately

If you have built a game, a meditation app, or a video editor with automatic sound design, you know the pain: you ask a sound generator for a three-action scene and get a muddle in which everything happens at once or half the events disappear. ElevenLabs has released SFX v3, and the update's main promise addresses exactly that: the model follows prompts containing multiple instructions more accurately.

What has changed

SFX v3 is a new model in the Sound Effects API. According to the company, it preserves sequences of multiple events better than v2, handles vocals and music more cleanly, and is less likely to clip high-frequency sounds such as gongs, bells, and a whistling kettle. ElevenLabs itself describes the release as incremental, so do not expect a revolution. This is a careful upgrade, not a new generation.

Migration is trivial for developers. In the text_to_sound_effects.convert call, change the model_id to eleven_text_to_sound_v3, and that is it. The endpoint remains the same: POST /v1/sound-generation. The announcement gives this example: “A stadium crowd cheers, a ball is kicked, the crowd boos.” That is three events in a single request, precisely the kind of scenario v3 is supposed to handle better.

Parameters worth knowing

The model selects the effect's duration automatically from the prompt text. If you need exact timing, set duration_seconds. If the result drifts away from the request, raise prompt_influence above its default value of 0.3: the higher it is, the more strictly the model follows the text and the less variation there is between generations. The output_format parameter sets the output format, such as mp3_44100_128 or pcm_48000.

One easy-to-forget limitation is that looping currently works only in eleven_text_to_sound_v2. If you are creating background loops, such as rain noise or the hum of a space station, stay on v2 with loop: true. This means that some projects cannot migrate completely to v3 and will need to retain both models.

How to write prompts

Order matters in the text. List events in the sequence in which they should be heard: “Footsteps on gravel, then a metal door opens.” For a single sound, a short and specific request works best.

Wording also affects the result. One ElevenLabs tester noted that “a couple arguing in French” (arguing) works noticeably better than “a couple fighting in French” (fighting). The practical takeaway is to try several synonymous descriptions of the same event before hard-coding the prompt.

What to do in practice

Because both models use the same endpoint, an A/B test is inexpensive. Take about ten of your most common production prompts, run them through v2 and v3, and compare them by ear. Pay particular attention to complex multi-part scenes and sounds with sharp high frequencies: that is where the improvement is expected. Keep v2 for loops.

Overall, this is a practical update for teams that have already integrated ElevenLabs SFX. There are few new capabilities, but changing one configuration line is worth a few minutes of testing. You can audition the results without code in the Sound Effects tab, in Flows, and in Studio within ElevenCreative.