Changelog

New visual detection release: V-JEPA 2 video backbone plus image accuracy gains and updated heatmap and tree-view analysis.

Voices created via the API now default to the rapid voice clone type.

Streamlined the voice creation flow and began deprecating all older voice models in favor of the current generation.

Detect now falls back to reading C2PA content credentials when present to support provenance-based verification.

Upgraded the Speech-to-Text backend to Gemini 3.5 for better transcription accuracy.

Added a dedicated report view summarizing watermark embedding and detection results.

Watermarking now supports both implicit (invisible) and explicit (custom message) marks.

New audio deepfake detection model deployed: equal error rate down to 4%, with higher precision, recall, and VoIP robustness.

Removed the Simulations and Agents entries from the left sidebar to streamline navigation.

The Guard product was renamed to Trace across the frontend, bundling Signal and Watermarking.

Removed the persistent "Unlock Voice Clones" upgrade banner from the app.

Added C2PA signing of outputs and SynthID detection to the watermarking stack.

Migrated Voice Design generation from Dreamvoice to the new Dramabox model.

Swapped the Intelligence expert model from Gemini 3.1 Pro to Gemini 3.5 Flash.

A dedicated playground for running and iterating on Intelligence analyses directly in the app.

Meetings now integrates natively into Microsoft Teams, with no bot joining the call and admins able to roll it out silently across the org.

Identity now distinguishes person and brand identity types and supports face matching via a Face Match API.

Moved audio watermarking to the PerTH V2 engine on Cerebrium for stronger, more robust marks.

Printed and exported deepfake reports now include a video thumbnail and the analysis heatmap.

Watermark detection now reads C2PA content credentials, with Resemble as an approved C2PA issuer.

A new Labs section surfaces models published under the research banner.

A new UI for applying and detecting watermarks across audio, image, and video.

Updated voice cloning so the first clone is free and each additional clone is $2, with card capture deferred until after the first clone.

Meetings now uploads each participant's stream to Intelligence post-call and shows the analysis.

Added support for AMR and 3GP audio formats across upload and processing.

Removed the legacy Fine-tuning entry from the left sidebar.

Intelligence reports now include Fraud Analysis, Liveness Detection, and Digital Alteration experts, all powered by Gemini.

New xAI-powered Ultra voice model now live in the portal, with six new default voices and $2 voice cloning (first clone free).

Rebuilt the Detect experience with a cleaner report layout and updated playground.

Voice Design previews now default to the Gemini 3.1 Flash TTS preview model.

Signal, an API-first intelligence layer on top of the detection stack, is now available to admins.

Detect report links are no longer publicly accessible - access is limited to team members and Resemble admins.

Added right-to-left (RTL) rendering and handling for transcripts in Speech-to-Text.

Refactored detection jobs to be provider-agnostic with a dedicated inference worker and Replicate polling fallback.

Voice cloning now supports non-speech source audio, including cloning a character voice from dog barks.

Admins can configure per-team Slack notifications that fire when a team runs a deepfake detection.

Added a Zero Retention mode so detection media is not persisted after analysis, behind a feature gate.

Speech-to-Text now supports files up to 4 hours long and 2GB in size.

Image and video detection model update improving robustness on content without faces, including war-related deepfakes.

Added an API endpoint to detect AI-generated text, extending Detect beyond audio, image, and video.

Enabled LoRA scaling in production and lowered the default CFG weight for more expressive cloned voices.

Detect can now run a reverse image search to add real-world context when assessing whether an image is authentic or AI-generated.

Define custom pronunciations per project for out-of-vocabulary words, with automatic cache sync and sub-10ms lookup at synthesis time.

Teams can now customize the display name of the Resemble meeting bot.

Admins can connect an org calendar so users and meetings auto-sync and bots auto-join based on preferences.

Added dedicated coordinator jobs and queues to handle long-form text synthesis reliably at scale.

Added the ability to delete multiple detections at once.

Added the Identity product into the main app experience.

Intelligence can now run multiple expert models and enable grounding for detect-driven analysis.

Switched the audio enhancement pipeline to use Cleanvoice AI as the primary provider.

The Meetings integration can now run deepfake voice detection on participant audio streams.

Updated the HiFi-GAN vocoder to eliminate high-pitch artifacts in synthesized audio.

Detect can now send email notifications when a detection completes.

Extended Intelligence reporting to images, adding analysis on top of image detection.