Introduction

I’ve known Karel since 2009, we met at eComm in San Francisco (surprisingly the event’s website sort of remains here, http://ecommconf.com/). He was a graduating MBA student from NYU Stern, in our discussion voice agents were mentioned even back then.

After graduation he joined Alcatel Mobile Phones, then moved to Orange Libon. During his time at Libon the idea about how voicemail could become a powerful assistant became a driving force in the creation of Voxist. And we’ve tracked his journey into voice AI, before voice AI was a thing.

Karel has been part of TADSummit and TADHack since the beginning. He was the first of us to point out the transformer model had changed the rules in AI.

One thing Karel said is “Flexibility is key when you’re working in AI.” And he’s been working with AI longer than most. Taking advantage of the weaknesses in French ASR (Automatic Speech Recognition) of the big companies he has built success mining voice data in verticals such as medical imaging. He’s thrived, but not yet made bank.

We’ll explore critical challenges facing the industry, and the recipes for success in Voice AI.

In the Beginning

eComm was a “who’s who” of innovators, Karel mentions the founder of Ring, Jamie Siminoff, was there. Jamie attended a private school close to me, Morristown Beard.

We use the term Voice AI, but it’s been around for decades with vocal assistants before Alexa and Siri. Karel included a demo of Serengi and WildFire from 1994 / 1996 in a previous session. Lots of attempts, and we’re slowly getting voice interfaces closer to mainstream.

Karel quotes the Latin proverb “Verba volant, scripta manent,” which means “spoken words fly away, written words remain.” Which simply, is what Voxist solves for businesses today. To do that they: transcribe voice, integrate into company data, and feed the robots to aid decision makers. Similar to VCONIC.

The Importance of Focus

The industries they have focused on include: telecom (voicemail assistants and call bots); AI transcribed medical reports (now at 10k reports per month); and a new area of implicit knowledge capture, indexing, and searching. Moving beyond RAG (Retrieval-Augmented Generation) into knowledge graphs.

Knowledge graphs are interconnected networks of data points that accurately model and represent data, down to individual values, and can incorporate trillions of nodes and relationships. Knowledge graphs have been around for years; major search engines and social networking sites use them to make the web more machine-readable for search, display, and advertising. In enterprises, knowledge graphs address challenges not fully addressed by cloud data platforms by integrating and querying complex data from diverse sources and formats.

The problem facing many corporations is baby boomers are retiring, hence corporates are losing tacit knowledge. Yes, there are stated processes, but human experience ensures those processes work. So Voxist has built the process to capture tacit knowledge from humans, through interviews.

Karel shared some Voice AI home truths. ASR is not perfect, because voice quality varies so much, models are less trained on female and child voices, as well as accents and english as a second language. The benchmarking can be ideal, that is studio quality voice, rather than mobile quality.

Sometimes a great ASR result is because of a massive hallucination failure. That is half the text was not spoken, rather made up by the LLM. Or the LLM does not generate an error, rather uses text that sounds like what is said, which on medical records can be dangerous.

Businesses have their own lingo, also brand names, street names, technical terminology needs specialist training. Real time can require significant GPU resources, which at some point will need to be closer to cost.

Voxist wins because of customer focus.

Mapping the Ecosystem

Karel then shares an interesting map of the Speech AI Ecosystem, see below. Capturing the many vendors, including new kid on the block Gradium. Karel highlights an inconvenient truth facing the industry on recognizing service credits to companies within investors’ portfolios. Many millions in recognized revenue without a sales force.

Bubble or a Growing Ecosystem?

Karel also shared how Nvidia and OpenAI fund the current money machine, source Bloomberg News. There is talk of OpenAI going IPO in 2026 with a 1T+ valuation. This situation is all based on confidence. We see insurance companies refusing to cover Gen AI risk. I remind everyone, Cisco survived the dotcom bust. But its valuation dropped by 80% and it took two decades to recover.

The current AI boom shares the “froth” and high valuations of the dot-com era. However, the underlying financial stability and real-world utility of the dominant companies suggest it may not result in a similarly catastrophic, economy-wide crash.

A potential correction might be more contained to specific, unprofitable AI start-ups rather than the broader tech market. Microsoft could buy OpenAI. There is always another DeepSeek reckoning, where racks of screaming nvidia boards are replaced with something much simpler and good enough. The approach Voxist takes to compete with OpenAI and make money could become widely known. Customer focus scales solutions down.

Karel brings up an interesting dilemma on the scope of a question to an AI, and the potential charge. Perhaps a postpaid model could be applied?

Karel also shared an experience of benchmarking their model versus OpenAI Whisper. Their model performed poorly compared to whisper on a data set. But the reality is OpenAI hallucinated for half the audio as the speaker had stopped talking and said nothing. In medical records, sometimes errors are good to highlight issues, than sweep them under the carpet with a slick hallucination.

Our consensus was AI will continue to do incredible things through the coming years. AGI is perhaps 5 years away. Voice AI agents will continue to improve and deliver great results, as long as they are focused on a specific businesses situation. We will likely mistake AGI though the many smart agents working together to give the perception of intelligence,

Karel raises the skills gap for humans in gaining the experience of a doctor analysing images for over 20 years. The same is true in software development, AI can do the work of interns, so they do not build the experience of a senior developer. Which we discussed in the AI Coding session.

Thank you to Karel, and well done on his progress in growing Voxist, and showing the many areas voice AI can help businesses.