ElevenLabs releases SFX v3, a sound effects model built for multi-sound prompts

ElevenLabs has released SFX v3, its newest sound effects model, available now in ElevenCreative and through ElevenAPI. According to ElevenLabs, the model follows prompts that describe several sounds more reliably, generates music and vocals more convincingly, and produces cleaner, more natural audio than SFX v2 in benchmark testing.

What changed from SFX v2

ElevenLabs describes three areas of improvement:

• Multi-sound prompts. SFX v3 is better at following several instructions within a single prompt, so more of the described sounds appear in the generated audio. ElevenLabs gives the example of a football clip that needs a crowd cheering, a ball dropping, and the same crowd booing seconds later. Until now, the company says, getting all of that from one prompt took trial and error.
• Music and vocals. Generation is improved for prompts that include music and vocals. ElevenLabs says this allows effects that pair a musical phrase or vocal performance with environmental and everyday sounds.
• Audio quality. The company says SFX v3 improves overall sound quality and is less likely to produce unwanted artifacts on sounds such as gongs, chimes, and whistles. ElevenLabs says the aim is to preserve a gong's resonance, a chime's ring, and a whistle's tone without harshness or crackle.

The benchmark claim comes from ElevenLabs. Benchmark scores and the test methodology have not yet been disclosed.

Where to use it

In ElevenCreative, users can create original sound effects from a text prompt in:

• Flows
• the Sound Effects tab
• Studio
• the Music Editor

SFX v3 is also available in the ElevenLabs mobile app. Developers can access it through ElevenAPI to add sound effects generation to their own products and workflows.

ElevenLabs suggests an iterative workflow for scene design: start with the main action, add the surrounding sounds, listen to how they combine, then adjust the description to move the result closer to the intended output.

Specifications

• Model: SFX v3
• Predecessor: SFX v2
• Availability: ElevenCreative, ElevenAPI, ElevenLabs mobile app
• Pricing: not yet disclosed
• Context window, parameter count, training cutoff: not yet disclosed

The announcement was written by Renee Zhang, Product Marketing at ElevenLabs.

What this means

The main change is in prompt adherence. Scenes with layered audio, such as a crowd reacting to a single event, previously required repeated generations to get every described element into the output, according to ElevenLabs. If SFX v3 delivers on that, a single prompt can cover a whole scene, which matters for creators working on video and games.

The addition of music and vocal handling to a sound effects model means these elements can be mixed with environmental sounds in one generation. The artifact reduction on tonal sounds like gongs, chimes, and whistles targets a specific quality issue that affects how believable an effect feels in a scene.

Because the model ships in ElevenAPI at launch, developers can integrate it directly. Teams evaluating it should run their own comparisons against SFX v2, since ElevenLabs has not published benchmark figures.