Meta Extends Its Segment Anything Line Into Sound With SAM Audio

What SAM Audio actually does

On December 16, Meta released SAM Audio, described as the first unified AI model that segments sound out of complex audio mixtures. The pitch is concrete: record a video of a band and isolate the guitar or vocals with one click, filter traffic noise out of footage shot outdoors, or strip a dog barking from an entire podcast recording.

The framing matters. Meta positions the model against a 'fragmented space' of single-purpose tools, where each editing task historically required its own specialized software. SAM Audio's claim is that one model handles the range of separation jobs people actually think about, rather than shipping a different tool per use case.

Three prompt types, one of them new

SAM Audio accepts three kinds of prompts. Text prompting lets you type 'dog barking' or 'singing voice' to extract a specific sound. Visual prompting lets you click on the person or object in a video that is making a sound, tying the audio track back to the visual frame. Span prompting lets you mark the time segments where the target audio occurs.

Meta calls span prompting an industry first. It is also the most engineering-relevant of the three, because it treats the timeline itself as the selection surface — the audio equivalent of drawing a bounding box over a region rather than describing its contents. Meta notes the three methods can be used alone or in any combination, which is where the 'unified' claim earns its weight: text, click, and time-range selections compose into a single separation operation.

The Segment Anything pattern, moved to a new modality

SAM Audio is explicitly framed as the latest addition to Meta's Segment Anything collection. The original SAM work established a prompt-driven interface for isolating regions in images and video; this release carries the same interaction model into audio, where the 'region' being segmented is a sound source inside a mix.

That lineage is the point. Meta is not just shipping an audio separator — it is extending a consistent prompting vocabulary across image, video, and now sound. Visual prompting bridges the two: clicking an object in a frame and getting its audio out treats the video and its soundtrack as a single addressable object.

We're already using SAM Audio to help build the next generation of creative media tools.Montana Labs

Availability and the open-model distribution choice

Meta made SAM Audio available two ways on day one: downloadable for direct use, and playable through the Segment Anything Playground, a new platform where anyone can select from Meta's audio and video assets or upload their own. The announcement carries an Open Source tag, consistent with how Meta has distributed earlier Segment Anything models.

The dual release — hosted playground plus downloadable weights — targets two distinct audiences at once: casual users testing the demo assets, and builders who want the model in their own pipelines. Meta also states plainly that it believes SAM Audio is the best audio separation model available, a benchmark claim the download release invites others to test.

What span prompting signals for audio tooling

The specific bet worth watching is span prompting. Text and click prompts are familiar from image segmentation, but marking time ranges as a selection primitive is native to audio and video work — it maps directly onto how editors already scrub timelines. If that interaction becomes a standard prompt type, audio editing tools built on SAM Audio could converge on a selection model that mirrors the cut-and-region workflows editors know, rather than the file-in-file-out batch tools that defined earlier source separation.

For teams building creative media software, the near-term implication is a single downloadable model that covers isolation, noise filtering, and source separation across the use cases Meta lists — music, podcasting, television, film, research, and accessibility — with a prompting interface that composes text, click, and time-range inputs instead of requiring a separate tool for each.