Resemble AI Detect voice and media deepfake detector
Vendor Resemble AI, Inc. Official site resemble.ai/products/detect How this card is sourced
Updated · AI Detector Index editors
Resemble AI Detect bills $0.035 per second of audio, $0.035 per image and $0.07 per second of video on its Flex plan, which has no subscription fee, per the vendor's pricing page (resemble.ai/pricing, read 9 November 2026). Detect is a voice deepfake detector that also scores images and video, not text: the product page says it "returns a verdict and an explanation for audio, video, and images", and the API accepts a file of up to 150 MB by direct upload (resemble.ai/products/detect and docs.resemble.ai/detect/create, read 9 November 2026). The vendor is Resemble AI, Inc. of Mountain View, California, which its own terms describe as a company that "produces synthetic voices and provides synthetic-media detection services", so the detector's maker also sells voice cloning. The card joins the detector index beside the other media cards, and how this card is sourced explains why a vendor's benchmark figures are quoted rather than ranked.
What Resemble AI Detect detects
| Field | Value | Source |
|---|---|---|
| Content types detected | Voice, image, video, music | resemble.ai/products/detect |
| Generators covered | Audio: ElevenLabs, OpenAI TTS (TTS-1, TTS-1 HD), Azure TTS, Google TTS (WaveNet, Neural2), AWS Polly, Suno, Udio, PlayHT, Murf, Descript Overdub, Bark, VALL-E, VALL-E 2, YourTTS, Tortoise TTS, XTTS, StyleTTS 2, MetaVoice, Kokoro, VoiceBox, NaturalSpeech 3, HierSpeech++ 'and over 100 more', Image: DALL-E 3 (98% accuracy), Midjourney v6 and v7 (98%), Stable Diffusion SDXL and SD 3.5 (94%), Nano Banana, Flux v2 (99%), Gemini Imagen 3 and 2.0 Flash (99%), GPT-4o image generation (99%), StyleGAN v2 and v3 (>99%), Ideogram v2 and v3, Leonardo AI, Latent Diffusion, eDiff-I, GLIDE, Kandinsky, DeepFloyd IF, Playground v3, Grok Aurora, Recraft v3, Lumina Image, Emu (Meta), Video: Sora, Veo (>99% (accuracy)), Runway Gen-3 Alpha and Gen-2, HeyGen, Pika, Kling, Seedance, Synthesia, D-ID, CogVideoX, Wan 2.1, Open-Sora, ModelScope, VideoCrafter, AnimateDiff, VideoPoet, Make-A-Video, Morph Studio, Mochi 1, HunyuanVideo, '250+ Generative AI models tested against in real time' with 'zero-day updates when new models launch' (product page) | resemble.ai/products/detect |
| Languages | Arabic, Chinese, Danish, Dutch, English, Finnish, French, German, Greek, Hebrew, Hindi, Italian, Japanese, Korean, Malay, Norwegian, Polish, Portuguese, Russian, Spanish, Swahili, Swedish, Turkish, Ukrainian, Vietnamese, Thai, Indonesian, Romanian, Czech, Slovak (54 listed) | resemble.ai/benchmarks |
| Input limits | API (POST app.resemble.ai/api/v2/detect, documentation read 9 November 2026): direct multipart upload up to 150 MB; a public HTTPS URL with no upload size limit; a secure upload token for files over 150 MB; a 200 MB source limit when detect_watermark is true. Allowed extensions: .wav, .mp3, .m4a, .ogg, .aac, .flac, .amr, .3gp, .3gpp, .mp4, .mov, .avi, .mkv, .webm, .jpg, .jpeg, .png, .gif, .webp; requests outside these constraints are rejected with HTTP 400 before billing. Audio and video are scored in windows of 1 to 4 seconds (frame_length, default 4), a segment can be selected with start_region and end_region, and max_video_secs caps processed video; a zero_retention_mode flag deletes the submitted file after detection. Batch and streaming (WebSocket) endpoints exist. Product page FAQ on formats: 'Audio: MP3, OGG, AAC, G.711, G.723.1, PCMu, PCMa. Video: all major formats from current AI generators.' Pricing: Team and Business plans add 'Large file uploads (150MB+)' and batch uploads; Flex bills $0.035 per second of audio, $0.035 per image and $0.07 per second of video. | docs.resemble.ai/detect/create |
| Free tier | Flex plan at $0 a month: 'no subscription fee', credits bought as you go and never expiring, with deepfake detection for audio, image and video, Intelligence insights, meetings detection and API access on 1 seat; the product page offers 'Try it free' and 'Get your first detection for free' with no stated allowance | resemble.ai/pricing |
| API | Yes, docs.resemble.ai/detect | docs.resemble.ai/detect |
| Integrations | REST API with Python, Node.js and JavaScript SDKs, batch and WebSocket streaming endpoints, MCP for Cursor and Claude Code, Deepfake Detection Chrome extension ('Detect deepfakes directly in your browser'), Microsoft Teams (featured partner, 'in just one click'), plus Zoom, Google Meet, Webex and phone systems for real-time meeting detection, On-prem, air-gapped and private VPC deployment (AWS, GCP, Azure); Docker and Kubernetes, Carahsoft (US federal, state and local government contract vehicles) | resemble.ai/products/detect |
| Data retention | Privacy Policy (last updated August 14, 2026), section 11: 'Audio, image and video submitted for detection analysis is deleted after analysis unless it is flagged. Meeting media ... of participants that are not flagged is analyzed in real time and discarded immediately afterwards; it is not stored. Where a participant is flagged, the recording of that participant is retained only for as long as the business customer that enabled the Services directs, and in any event no longer than the period set out in Section 3, and is deleted on that customer's instruction. Detection results and associated metadata are retained for the duration of the business customer's account, or as that customer directs.' Section 3: 'Biometric identifiers are permanently destroyed when the initial purpose for collecting them has been satisfied, and in any event no later than three (3) years after they were collected, whichever occurs first.' Deletion requests are completed 'within thirty (30) days'. Product page: 'Zero Retention Mode ensures submitted media is permanently deleted after analysis and never viewable by Resemble staff or systems'; the API's zero_retention_mode flag 'automatically delete[s] the submitted media file after detection completes' and redacts media URLs from responses. | resemble.ai/privacy-policy |
| Ownership | Resemble AI, Inc., 812 W Dana St, Mountain View, CA 94041, United States (Privacy Policy and Terms of Service, both last updated August 14, 2026); the terms are governed by Delaware law with arbitration seated in Wilmington, Delaware. The terms describe the company as one that 'produces synthetic voices and provides synthetic-media detection services', so the detector's vendor also sells voice cloning. Services are 'hosted and operated in the United States through Resemble AI and its service providers, principally Google Cloud Platform (us-east4, Virginia), Render (Oregon) and Cerebrium (United States)'. The current model is DETECT-World, which 'expands on DETECT-3B Omni's foundation'; earlier models are DETECT-3B Omni and DETECT-2B. The company lists Carahsoft public sector contracts and a US Senate Judiciary testimony on its product page. | resemble.ai/terms-of-service |
The current model is DETECT-World, which the product page describes as "the first world model for deepfake detection", one that reviews "both known generation patterns and physical reality, including lighting, motion, and continuity, so it can flag content from generators it hasn't been trained on" (resemble.ai/products/detect, read 9 November 2026). The page's coverage table runs to "250+ generative AI models". For audio it lists ElevenLabs, OpenAI TTS, Azure TTS, Google TTS, AWS Polly, Suno, Udio, PlayHT, Murf and Descript, then Bark, VALL-E, XTTS, StyleTTS 2 and others "and over 100 more"; for images DALL-E 3, Midjourney v6 and v7, Stable Diffusion SDXL and SD 3.5, Nano Banana, Flux v2, Imagen 3, GPT-4o image generation, StyleGAN, Ideogram and Leonardo, each with a stated detection figure between 94% and over 99%; for video Sora, Veo, Runway, HeyGen, Pika, Kling, Seedance, Synthesia and D-ID. Suno and Udio make music a fourth modality in the vendor's own list.
An attack coverage table marks voice clone, replay attack, face swap, partial edit, synthetic image, live call interception and video conferencing as "DETECTED", under the heading "Can it catch a fake it's never seen before? That's why we test and update continuously." Every detection passes through a layer the vendor calls Resemble Intelligence, which "generates human-readable forensic explanations in real time" and is "Powered by Gemini 3 Flash", and a reverse image search "Built to reduce false positives" that checks submitted images against live web sources (resemble.ai/products/detect, read 9 November 2026). Language is a property of the audio model: the benchmarks page states "54 languages detected from Arabic to Spanish", names 30 of them, and says detection "relies on generation artifact patterns, not language-specific features" (resemble.ai/benchmarks, read 9 November 2026).
The company is Resemble AI, Inc., 812 W Dana St, Mountain View, CA 94041, under a Privacy Policy and Terms of Service both last updated August 14, 2026, with Delaware law and arbitration seated in Wilmington (resemble.ai/terms-of-service, read 9 November 2026). The services are hosted "principally" on Google Cloud Platform in Virginia, Render in Oregon and Cerebrium in the United States. The sibling voice card is Pindrop Pulse, whose own figure appears on Resemble's benchmarks page; the sibling media cards are Sensity and Reality Defender, and the two image cards with a study record are Hive and AI or Not.
Limits and plans
| Plan | Price | Limits | Checked on |
|---|---|---|---|
| Flex | $0 /mo | No subscription fee; prepaid credits; Detect audio $0.03500 per second, images $0.03500 per image, video $0.07000 per second; Intelligence $0.02500 per second; 1 team seat; platform and API access; real-time detection for meetings; SynthID check and C2PA signing. | |
| Team | $350 /mo billed monthly, or $280 /mo billed annually (save $840/yr) | Everything on Flex; Detect audio $0.01500 per second, images $0.01500 per image, video $0.03000 per second; Intelligence for meetings included; large file uploads (150MB+); batch uploads; 5 team seats. | |
| Business | $1,000 /mo billed monthly, or $800 /mo billed annually (save $2,400/yr) | Everything on Team at the same processing rates; configurable as deployment grows; org-wide calendar integration for Meetings; Meetings security notifications; Single Sign-On (SSO); 20 team seats. | |
| Enterprise | Contact us: volume pricing, enterprise SLAs, custom model training, on-premises deployment, dedicated support, SOC type 2 documentation; the FAQ says a Business customer spending over $1,000 a month 'could save significantly with volume pricing'. |
The API documentation is the source for limits, read 9 November 2026: a direct multipart upload of up to 150 MB, a public HTTPS URL with "No upload size limit", and a secure upload token for files over 150 MB; a 200 MB source limit applies when watermark detection is switched on. Allowed extensions are .wav, .mp3, .m4a, .ogg, .aac, .flac, .amr, .3gp, .3gpp, .mp4, .mov, .avi, .mkv, .webm, .jpg, .jpeg, .png, .gif and .webp, and "Requests outside these constraints are rejected with HTTP 400 before any analysis or billing runs" (docs.resemble.ai/detect/create). Audio and video are scored in windows of 1 to 4 seconds, a segment can be selected, video duration can be capped, and a zero_retention_mode flag deletes the submitted file after detection. The product page FAQ lists audio formats as "MP3, OGG, AAC, G.711, G.723.1, PCMu, PCMa", the telephony codecs its call integrations need.
Pricing is pay as you go with optional seats. Flex is "$0 /mo", "For builders and teams exploring detection with no subscription fee", with credits that "never expire"; Team and Business add seats and halve the per-second rates; Enterprise is custom (resemble.ai/pricing, read 9 November 2026). The product page offers "Try it free" and "Get your first detection for free" without a stated allowance, so the free tier field records the Flex plan and no word figure. Figures below are as displayed on 9 November 2026 in US dollars.
Retention is conditional on the verdict. The Privacy Policy of August 14, 2026 states: "Audio, image and video submitted for detection analysis is deleted after analysis unless it is flagged", that unflagged meeting media "is analyzed in real time and discarded immediately afterwards; it is not stored", and that "Where a participant is flagged, the recording of that participant is retained only for as long as the business customer that enabled the Services directs, and in any event no longer than the period set out in Section 3", which destroys biometric identifiers "no later than three (3) years after they were collected" (resemble.ai/privacy-policy, read 9 November 2026). Detection results and metadata "are retained for the duration of the business customer's account". The card's retention rank field takes the three-year cap on flagged media, 1,095 days, as the longest stated period, and quotes the delete-after-analysis sentence beside it; a Zero Retention Mode, on the product page and as an API flag, "ensures submitted media is permanently deleted after analysis and never viewable by Resemble staff or systems".
What the vendor claims
Product page: DETECT-World 'Up to 99.5% (accuracy)', 'Zero day coverage', '250+ models'; '#1 Ranked on third-party benchmarks for audio deepfake detection'. Benchmarks page (last updated August 2026): '99.5% Pooled detection accuracy, Podonos Audio DFD Bench' with '<1% False positive and false negative rate'; audio false negative rate 0.4% on 'Test sets include mp3, wav, flac, ogg, m4a, webm across 4,524 files'; images '95.8% Average accuracy ... when benchmarked against 1,000 image files'; video '98.2% Average accuracy ... when benchmarked against 1,000 video files'; the same page lists competitors' Podonos figures (Whispeak 97.7%, Aurigin AI 96.8%, Pella Research 95.8%, Pindrop 95%). Per-generator image figures on the product page run from 94% (Stable Diffusion SDXL and SD 3.5) to over 99% (StyleGAN). The Podonos benchmark is not an index study record.
The headline is "Up to 99.5% (accuracy)" for DETECT-World, with "Zero day coverage" and "#1 Ranked on third-party benchmarks for audio deepfake detection" (resemble.ai/products/detect, read 9 November 2026). The benchmarks page, last updated August 2026, attributes the figure: "99.5% Pooled detection accuracy, Podonos Audio DFD Bench", with "<1% False positive and false negative rate" on the same benchmark and a 0.4% false negative rate on test sets of "mp3, wav, flac, ogg, m4a, webm across 4,524 files" (resemble.ai/benchmarks, read 9 November 2026). For images the page states "95.8% Average accuracy" with a "False negative rate of 4.3% and false positive rate of 4% when benchmarked against 1,000 image files", and for video "98.2% Average accuracy" with "False negative rate of 2% and false positive rate of 0.95% when benchmarked against 1,000 video files".
The same page prints competitors' Podonos figures, Whispeak 97.7%, Aurigin AI 96.8%, Pella Research 95.8% and Pindrop 95%, and a false negative chart with Reality Defender at 3.6% and Pindrop at 3.7%. The Pindrop Pulse card quotes Pindrop's own account of the same benchmark, which calls Pindrop "the fastest most accurate commercial detector evaluated"; the index adopts neither vendor's figure for the other, and the Podonos benchmark is not an index record. The vendor's caveats sit at the foot of the benchmarks page: "EER and accuracy figures from the Hugging Face Speech DF Arena and DFBench Speech/Image leaderboards, March 2026. Resemble AI does not control test set composition. Image figures from DFBench Image 2025. Video EER from internal evaluations on held-out test sets", and "PerTh V2 figures are pre-release internal evaluations" (resemble.ai/benchmarks, read 9 November 2026). The API documentation adds that the attack classification has "no confidence cut-off: the first entry is always the best guess" (docs.resemble.ai/detect, read 9 November 2026).
A vendor that sells voice cloning and voice-clone detection from the same site has a reason to be good at both and a reason to be read carefully on both; the card records the claims with their stated bases and leaves the ranking to figures an outside paper prints.
What independent studies found
No record in the studies index tested Resemble AI Detect, read 9 November 2026 in the data file, and no indexed paper has scored any voice detector; the one indexed media record, Ha and Passananti 2024, scored Hive, Optic and Illuminarty on images and did not include Resemble. The study count is zero and the false positive, false negative, robustness and claim gap fields are empty.
The benchmarks the vendor cites, the Podonos Audio DFD Bench, the Hugging Face Speech DF Arena, DFBench and MLAADv10, are public leaderboards and datasets rather than papers the index has read, and the vendor says it "does not control test set composition". Each is a candidate for a study record; until one is read in the primary source, the 99.5%, 95.8% and 98.2% stand as the vendor's figures with the vendor's stated bases.
Where Resemble AI Detect appears in the index
Rankings
- AIDI-R-04AI detectors ranked by languages supported
- AIDI-R-07AI image, voice, video and code detectors ranked by independent studies
- AIDI-R-11AI detectors ranked by named integration surfaces
- AIDI-R-12AI detectors ranked by the upload formats the vendor states
- AIDI-R-17AI detectors ranked by the generator models the vendor names
Resemble ranks in two lists, read 9 November 2026 in the data file: AI detectors ranked by data retention, on the stated 1,095-day cap for flagged media, with the delete-after-analysis clause quoted in its row; and AI detectors ranked by language count, on the 54 languages the benchmarks page states, 30 of them named. It is absent from the study-based rankings, which need an indexed record; from the free tier ranking, which sorts on words and finds seconds and images here; and from the claim versus evidence gap, because no study pairs with the 99.5%.
The card appears on AI detectors by content type, where voice, image, video and music are its columns and text is blank, on detectors with an API for the documented v2 detect endpoint, on platform integrations for its Chrome extension, MCP server and REST API, on accepted file types for the 17 counted extensions, and in studies by modality with zero records against four non-text types. Every detector card on the register shows the same fourteen fields, so a per-second media price can be read beside the per-word credits of the text tools.
Resemble Detect's on-prem, air-gapped and private VPC deployment places it among the fourteen cards on the offline question, added 9 November 2026, and its 62 named generators place it first on the generators covered ranking, added the same day, where every count is the vendor's own list.
Questions about Resemble AI Detect
How much does Resemble AI Detect cost?
Flex has no subscription fee and bills $0.035 per second of audio, $0.035 per image and $0.07 per second of video; Team is $350 a month ($280 billed annually) and Business $1,000 ($800 annually) at half those rates, read 9 November 2026 on resemble.ai/pricing.
How accurate does Resemble say Detect is?
"Up to 99.5% (accuracy)" on the Podonos Audio DFD Bench with under 1% false positives and false negatives, 95.8% on 1,000 image files and 98.2% on 1,000 video files, per the benchmarks page last updated August 2026 and read 9 November 2026. No indexed study tested it.
What file sizes and formats does Resemble Detect accept?
Direct uploads up to 150 MB, URLs with no size limit and a secure upload path for larger files, in .wav, .mp3, .m4a, .ogg, .aac, .flac, .amr, .3gp, .3gpp, .mp4, .mov, .avi, .mkv, .webm, .jpg, .jpeg, .png, .gif and .webp, per the API documentation read 9 November 2026.
Does Resemble keep submitted audio and video?
Media is "deleted after analysis unless it is flagged"; flagged recordings are kept as the business customer directs and no longer than three years, and a Zero Retention Mode deletes every submission after detection, per the Privacy Policy of August 14, 2026 read 9 November 2026.
Does Resemble AI also make deepfakes?
Its terms describe a company that "produces synthetic voices and provides synthetic-media detection services", so the vendor sells voice cloning and voice-clone detection from the same site, read 9 November 2026.
Sources and revisions
Sources
- resemble.ai/products/detectRead 9 November 2026
- resemble.ai/pricingRead 9 November 2026
- resemble.ai/benchmarksRead 9 November 2026
- resemble.ai/privacy-policyRead 9 November 2026
- resemble.ai/terms-of-serviceRead 9 November 2026
- docs.resemble.ai/detectRead 9 November 2026
- docs.resemble.ai/detect/createRead 9 November 2026
- docs.resemble.ai/detect/secure-uploadsRead 9 November 2026
Revision log
- Card created2026-11-09: offline question and generators covered ranking linked (seventh wave)
