resemble-detect

What it does

Analyze audio, image, video, and text for synthetic manipulation, AI-generated content, watermarks, speaker identity, and media intelligence using the Resemble AI platform.

How it works

• It instructs to perform deepfake detection by submitting media to endpoints like POST /detect and then poll GET /detect/{uuid} for an asynchronous result labeled as either "completed" or "failed" with a corresponding score, label, and status. If a detection is completed, results include per-type metrics (audio, image, video) and may include fields such as metrics, image_metrics, or video_metrics with a label and score and additional detail like aggregated_score, consistency, and heatmap URLs.
• It supports combined analysis by using an intelligence flag in the same detect call to run multimodal analysis alongside detection.
• It can perform Audio Source Tracing when audio_source_tracing is set to true in POST /detect, yielding a label such as elevenlabs or resemble_ai when audio is detected as fake.
• It can apply or detect watermarks via POST /watermark/apply and POST /watermark/detect, returning a watermarked_media URL or { "has_watermark": true, ... } responses depending on the endpoint.
• It offers standalone Intelligence via POST /intelligence to extract structured fields such as speaker_info, language, emotion, transcription, misinformation, and image-specific fields like scene_description, authenticity_analysis, and misinformation.
• It enables Identity verification (beta) with POST /identity to create profiles and POST /identity/search to compare unknown audio against known profiles, returning matches with uuid, name, confidence, and distance.
• It includes a Text Detection module (beta) via POST /text_detect to determine whether text is AI-generated or human-written, with options to wait synchronously or poll asynchronously.

When to use it

Use this skill when the user needs: deepfake detection across media types, authenticity verification, provenance via source tracing, watermarking for provenance, speaker identity verification, or media intelligence (emotion, transcription, misinformation). It is not to be used for text-to-speech generation, voice cloning, or speech-to-text transcription, which are covered by separate capabilities.

What it can touch

• API endpoints: POST /detect, GET /detect/{uuid}, POST /watermark/apply, POST /watermark/detect, POST /intelligence, POST /audio_source_tracings (standalone), POST /identity, POST /identity/search, POST /text_detect, POST /detects/{detect_uuid}/intelligence
• Media URLs must be publicly accessible over HTTPS; local files must be hosted publicly first.

Caveats

• The system requires a valid Resemble API key and base URL https://app.resemble.ai/api/v2 with Authorization: Bearer <RESEMBLE_API_KEY>.
• Results must be completed (status: "completed") before making claims about authenticity; if status is "processing" or "failed", no autonomous judgment is allowed.
• Text and some features (Intelligence, Text Detection, Identity) are Beta features and may require special access or billing plans.
• The skill advises returning explicit scores and labels (e.g., label and score) as part of results and not making authenticity claims without a completed detection result.