Pindrop Security Inc. today announced the launch of its Pindrop Pulse Inspect deepfake detection tool in preview mode, saying it can detect artificial intelligence-generated speech in any digital audio file with 99% accuracy.

A deepfake is synthetic media in which a person in an existing image, audio or video recording is replaced with false content. Recent advances in AI have enabled deepfake creators to generate highly realistic yet entirely fictitious visuals and audio at little or no cost, with some tools available under open-source licenses.

There’s no federal law against them, although some legislation has been passed at the state level. Deep Media Inc. has estimated about a half-million deep fake video and audio recordings were shared on social media last year.

Pindrop has raised more than $318 million, including a $100 million funding round last month. Until now, it focused mainly on contact centers, where scammers can use synthesized voice to bypass biometric authentication systems, change profile data and gather information about account holders.

The company says 11 of the largest insurers and eight of the top 10 banks and credit unions in the U.S. are customers. It has detected more tha 104 million fraudulent calls, saving its customers $2 billion.

Pulse Inspect uses neural networks trained on more than 350 deepfake generation tools, 20 million unique utterances and more than 40 languages covering over 90% of the internet’s spoken languages.

Users can upload audio files into the application program interface-driven batch-processing platform, customized with adjustable thresholds based on risk tolerance. They quickly get a score that indicates the likelihood that the file contains synthetic speech, with pointers to specific sections that merit further analysis. Continuous deepfake scores are provided every four seconds.

Satisfaction guaranteed

Pindrop expanded into generalized deepfake detection earlier this year with the launch of Pindrop Pulse, a tool that analyzes audio to identify fake videos through voice recognition. It promoted its accuracy claims by offering a warranty compensating eligible customers if the tools fail to detect a deepfake or synthetic voice fraud.

With this release, the company is expanding its market to include media organizations, nonprofits, government agencies and social media networks. It’s primarily targeting business customers but may release a consumer detection tool in the future, Chief Executive Vijay Balasubramaniyan told VentureBeat. A Pindrop report released last fall estimated that more than 12 million U.S. adults know someone who has personally had deepfakes of them created without their consent.

Pindrop said it has had notable invisible recent success, including identifying the source text-to-speed engine used in a Biden deepfake incident earlier this year and a more recent case involving a Kamala Harris video parody last month.

Image: Pixabay