# Meta introduced SAM Audio with text, visual and temporal prompts for source separation Meta introduced SAM Audio on December 16, 2025, pairing multimodal audio separation with PE-AV, an evaluation benchmark and a reference-free judge model. Canonical: https://rangoon.ai/insights/archive-meta-sam-audio-december-2025/ Author: Rangoon Editorial (https://rangoon.ai/insights/editorial/) Published: 2026-10-24 Updated: 2026-10-24 Event date: 2025-12-16 Tags: [Meta AI](https://rangoon.ai/insights/tags/meta-ai/), [SAM Audio](https://rangoon.ai/insights/tags/sam-audio/), [audio AI](https://rangoon.ai/insights/tags/audio-ai/), [multimodal models](https://rangoon.ai/insights/tags/multimodal-models/) Source note: Historical reporting: Meta introduced SAM Audio on December 16, 2025. Its launch post and same-day research publication describe a promptable audio-separation model, PE-AV, SAM Audio-Bench and SAM Audio Judge. Reported state-of-the-art results are Meta research claims under stated tests; the event is not proof of complete separation, rights clearance, safety, accessibility outcomes or production suitability for a particular recording. Image: Prompted audio separation. Original Rangoon explanatory diagram: Source boundary: dated primary release reporting; original editorial analysis. No vendor interface, photograph, measurement result, or endorsement implied. ## Key takeaways - SAM Audio combined language, visual and time-span prompts in one source-separation model family. - Meta paired the model with PE-AV and two evaluation artifacts, separating generation, benchmark data and judging functions. - Research results for controlled tests do not guarantee clean isolation, ownership clearance or correct output in another recording or workflow. ## The December release framed audio separation as a multimodal prompting task [source 1](https://ai.meta.com/blog/sam-audio/) [source 2](https://ai.meta.com/research/publications/sam-audio-segment-anything-in-audio/) On December 16, 2025, Meta introduced SAM Audio as a unified model for separating sound from mixed audio or audiovisual material. The launch described three prompt paths: text that names a target sound, a visual cue on a sounding object or person in video, and a temporal span marking where target audio occurs. Those interfaces make the task more specific than asking a model to improve an entire recording. They do not establish that a target is present, that the prompt has only one reasonable interpretation or that the separated track is complete. Meta said it was releasing SAM Audio with Perception Encoder Audiovisual, or PE-AV, along with SAM Audio-Bench and SAM Audio Judge. The launch record makes those distinct components visible on the same date. A separation model produces a target and residual track, an audiovisual encoder supplies representation features, a benchmark supplies test material, and a judge estimates output quality. Bundling them in one announcement does not mean each component answers the same question or that a judge score replaces human review of an important recording. ## Meta described a flow-matching design and an audiovisual representation layer [source 1](https://ai.meta.com/blog/sam-audio/) The launch post describes SAM Audio as a flow-matching diffusion-transformer system that takes a mixture and prompts, maps them into a shared representation and generates target and residual audio. It says the training-data engine combines audio mixing, automated multimodal prompt generation and pseudo-labeling, with real and synthetic mixtures across speech, music and general sounds. These are technical design claims from Meta’s release, not evidence that every unseen sound can be separated without artifacts, leakage or ambiguity. ## The research record made benchmark scores conditional on its stated evaluation setup [source 1](https://ai.meta.com/blog/sam-audio/) [source 2](https://ai.meta.com/research/publications/sam-audio-segment-anything-in-audio/) Meta’s same-day research publication says SAM Audio was trained on large-scale multimodal mixtures and evaluated on general sound, speech, music and musical-instrument separation. It reports state-of-the-art results across a set of benchmarks and introduces a real-world benchmark with human-labeled multimodal prompts. Those are useful research claims because they name tasks and evaluation assets. They remain conditional on the benchmark composition, prompts, reference material and scoring protocol, rather than predicting quality for a broadcaster, archive, call centre or assistive device. SAM Audio Judge is presented as a reference-free evaluator whose output was designed to correlate with human judgment. A reference-free method is useful when isolated source tracks are unavailable, but correlation is not identity with every listener’s assessment or a legal finding about a recording. An automatic score can guide comparison while a user still needs to inspect errors, protect sensitive audio and decide whether an edit changes meaning. The December event released an evaluation method; it did not certify content decisions made from its scores. ## The research release did not promise universal audio cleanup or rights outcomes [source 1](https://ai.meta.com/blog/sam-audio/) [source 2](https://ai.meta.com/research/publications/sam-audio-segment-anything-in-audio/) Meta’s post gives examples such as isolating an instrument or reducing traffic noise, but examples demonstrate an intended use rather than a guarantee. Separation can be difficult where sources overlap strongly, a target is faint, the recording is degraded or a prompt covers more than one event. The release materials should also not be read as permission to process or republish an audio source. Rights, consent, privacy and distribution choices remain separate from a model’s technical ability to generate a new track. ## Sources 1. [Meta AI: Introducing SAM Audio: The First Unified Multimodal Model for Audio Separation](https://ai.meta.com/blog/sam-audio/) 2. [Meta AI Research: SAM Audio: Segment Anything in Audio](https://ai.meta.com/research/publications/sam-audio-segment-anything-in-audio/) ## Related reading - [Suno released v4.5 with longer songs and new prompt controls](https://rangoon.ai/insights/archive-suno-v4-5-may-2025/) - [Stability AI introduced Stable Audio 2.5 with ARC post-training and inpainting](https://rangoon.ai/insights/archive-stability-audio-2-5-september-2025/) - [Stability AI and Arm released Stable Audio Open Small](https://rangoon.ai/insights/archive-stability-audio-open-small-may-2025/) ## More information - [Documentation index](https://rangoon.ai/llms.txt) - [AI access and policies](https://rangoon.ai/ai/) - [Source repository](https://github.com/hypler-dev/rangoon)