Slavko Kovačević

Today we launch @deepmark_me (@ycombinator F26). We put an inaudible ID inside an AI agent's voice, so the person on the other end of a phone call can verify who is really speaking. Here is me doing it live, on one phone.

I would be lying if I said I felt successful. We are far, far from it. We got into YC, true, but accelerators and investments don’t define success. Truth be told, we are working with some amazing companies, and the feeling that our tech is being used and recognized by them is really great. But we are also proud of how far we’ve come. DeepMark was born in a post-communist country that barely has any startups. We were born into working-class families that considered stable government jobs the pinnacle of success. The university we worked at wasn’t even able to provide us with work computers for our research. Hell, Kosta and I had to buy Google Drive storage and Colab on the Romanian black market, as the legit ones were too expensive for us! And the scientific work we produced outperforms top companies and their research teams 5 times our size. Still, for a company to succeed, much more needs to align apart from a great product. And we are going to prove to the world and to ourselves that hard work pays off. We are not stopping until the word “DeepMarked” becomes synonymous with digital trust.

We built ANNI because the field still grades watermarks with PESQ, a 25-year-old metric made to score phone calls, not to tell you if someone can actually hear a watermark. ANNI compares two recordings by what the ear actually sends to the brain. On the ODAQ benchmark from Fraunhofer IIS and Netflix, which tests 38 quality metrics against real listeners, the gap between ANNI and second place is bigger than the gap between second and tenth, so this is a huge step forward. It's open source and already used by engineers from Deepgram, Netflix, and ElevenLabs! deepmark.me/blog/goodbye-pes…

Imagine Fat Boy and Ron ganging up on you with voice clones of your boss. You wouldn't stand a chance. $129M gone. nbcnews.com/news/us-news/3-f…

If I had a dollar for every "Hi Slavko, a watermark that survives phone calls..." LinkedIn message I received, I wouldn't need a seed round

I should have picked a slightly more vague picture, but do read the blog. You might learn a thing or two about watermarking, and there are like two decent jokes bundled within it, on the house!

The startup world loves e^x growth. These are my citations in the field of watermarking. What do you think? Is it taking off?

This is a massive L take. We open-source most of our research, as we love sharing knowledge. But we deserve to be credited if someone uses it. I would be pretty mad if someone wrote a prompt, "audio watermarking Claude help pls" and it spat out our proprietary models.

Starting November 7, xAI will add text watermarking to Grok API models by default. This means every major LLM is now watermarked.

OpenAI researched text watermarking for years but didn’t put it into production. When compliance deadlines were near they let Anthropic go first and take all the hate, haha. You gotta respect that, very smart 😆

Ladies and gentlemen, a new record has been set in the AI voice fraud market. One hacked voice and $108M departed Italy's biggest bank. Poof! reuters.com/legal/government…

The title is my takeaway from a new study by security researchers at Brno University of Technology, published in August. They tested how well people catch AI voices by ear. Spoiler, it doesn’t go well!

Not if you ask us.

stop the nukes as long as there is even one soul reading dostoevsky in sf

Our models run in your infrastructure. Your audio stays there. Please keep the hackers in your thoughts during this difficult time.

hey @grok can you tell me who pioneered deep-learning-based audio watermarking

I tested @ElevenLabs, @cartesia, @FishAudio, and @MiniMax_AI voice cloning capabilities. The results are amazing!

Only in SF as they say

I wrote about voice cloning and what happens when even our real voices become harder to believe.