Neural Defend ranked #1 in the world for audio deepfake detection

VMPL
New Delhi [India], November 12: Neural Defend has been ranked number one in the world for detecting AI-generated voices. On an independent, publicly scored benchmark run by voice-AI evaluation firm Podonos, its Arcas 290M and Arcas 90M models hold first and second place among 28 detection systems, each scoring 99.93% on 4,524 audio clips.
The result
Each model made three errors in 4,524 files (false-positive rate 0.1%, false-negative rate 0.0%). Podonos called it “the closest any system has come to a clean sweep of this set.” Of the 28 systems on the leaderboard, 18 are commercial products from 16 vendors (the rest are open-source and research models), among them the best-funded names in the category: Pindrop ($100 million in financing in 2024) ranked 13th at 95.05%, Hive (valued at $2 billion, with a US Department of Defense deepfake-detection contract) 17th at 83.53%, and Reality Defender (now supplying Orange Business) 18th at 71.27%.

The smaller Arcas 90M matches the top score, runs a quarter faster at 89 ms per file, and is small enough to run on a phone with no internet connection.
Why it matters
Most vendors publish accuracy figures measured on test sets they control, so the numbers cannot be compared. Podonos, an independent evaluator backed by NAVER, keeps the answer keys private and scores every vendor’s submitted predictions itself, on the same clips, with the same method. The leaderboard is public at podonos.com; Podonos founder Soohyun Bae also congratulated Neural Defend publicly on LinkedIn.
Quotes
“We sent Podonos our predictions without knowing the answers, and we came out first,” said Sivashankar Selvarajan, CTO of Neural Defend. “Every vendor claims 99%. This is the first time the whole market has been scored on the same files by someone with no stake in the result.”
“The part we are proudest of is that our 90M model performs as well as much larger models from unicorn companies, and solves the problem at a fraction of their cost,” said Sumit Singh, CDO of Neural Defend.
“A cloned voice now costs almost nothing to make; we intend to make it cost just as little to catch,” said Piyush Verma, CEO of Neural Defend. “AI voice fraud is growing fast and it is global: banks, contact centres, insurers and telecoms are all facing it. Our aim is to be the state of the art in detecting it, and to get that protection into their hands quickly.”
Availability
The same accuracy comes two ways:  Arcas 90M on device , fully offline, so audio never leaves the handset or the customer’s network; and  Arcas 290M in the cloud , via API or SDK, on-premises or in any cloud the customer chooses. Neural Defend is deploying it wherever a voice is trusted to prove identity or authorise an action: banking and payments, call centres, KYC, insurance, ride and delivery platforms, matrimony and dating, and media, with emerging demand from meeting and contact-centre platforms, telecom networks, e-signature services, voice-AI platforms and government.
About Neural Defend
Neural Defend builds AI deepfake detection for voice, face and video, deployed on device, on-premises, in a private cloud or through an API, and used to stop cloned voices and synthetic faces in onboarding, authentication, payments and customer support. Headquartered in New Delhi, India, it works with clients across India, the Middle East, Southeast Asia and the United States. neuraldefend.com
(ADVERTORIAL DISCLAIMER: The above press release has been provided by VMPL. ANI will not be responsible in any way for the content of the same.)