BotCast #9

Meet Katrin Tomanek and Alex Kauffmann, Co-founders at Squeezlabs, who joined the BotCast #9 to answer our usual questions on building with Voice AI.

Who are you?

Katrin: My name is Kat. I am an AI researcher. I think I've worked on the whole stack and mostly focused on automatic speech recognition. Together with Alex, we're building small and local AI solutions.

Alex: I'm Alex. I worked in R&D for a long time and met Kat at an incubator, where we both bonded over our worry over privacy and our like of small, reasonable things over gigantic SUVs. We've been working on a company called Squeezlabs for the last 18 months, and have built some satirical things and some serious things.

How did you end up using voice AI?

Katrin: I worked in a project in Google Research on a long-tail speech pattern that was about making general speech recognition work for people with non-standard speech. That got me really deep into understanding how ASR works, where it doesn't work, and what you can do to actually make it work for everyone instead of just a few people.

Alex: My predominant interest was in local models and voice was one of the modalities, but Kat was really interested in voice models and I was really interested in working with Kat, so I got into voice models.

A moment of joy and pain using voice tech.

Alex: My biggest pain actually is watching my six-year-old son try to interact with my Google TV. He is constantly asking for cartoons, they are Spanish cartoons, so he is asking in Spanish and he has got a little child's voice and he gets all kinds of results, never the results he wants, and he gets so frustrated. While it is very fun for me to watch, I do think it is a clear failure point for Voice AI.

Katrin: My joy connects there because I have worked with a lot of people for whom ASR also doesn't work because of how they speak. I know from experience that you can change that, you can adapt models and even small models by specializing them to someone's specific speech patterns and all of a sudden they will understand you, and that can be massively life-changing for people who before have never had a chance to be understood.

What lesson would you share with other builders?

Alex: General intelligence and large models are principally for people who don't know what they want to build. We think that if you have a specific use in mind and you have tailored that use and the constraints and have a product that you're actually trying to build with it, it's much easier to define exactly what your needs are in such a way that you can rely on small, local models that will run on the hardware that you have, rather than sending your data to god knows what cloud and tying yourself in to sort of open-ended, uncapped costs and the like.

Where do you see the industry in 12 months and in 5 years?

Katrin: I believe voice AI will be everywhere, it will become some of the standard interfaces, and that is on everyone's mind right now. But the hope would be that in addition to that, a lot of this voice processing is handled locally and will be able to understand you specifically, as opposed to bringing all your data in the cloud with all the privacy constraints, with all the connectivity constraints it has, and all the potential costs.

Alex: Where I see it and where I would like to see it are two different things. I unfortunately have a very dystopian view of what is going to happen if we are not careful. I think it is already happening in China, for instance, where just having everything listening and everything transcribing means that it is almost free to have complete surveillance of everything, which for the most part is probably benign, but it just opens up possibilities that we have been watching in bad sci-fi since we were kids. What I would really like to see, especially on the longer term, is that voice moves from being part of the software stack to actually being a part of hardware, where voice models get good enough and are well understood enough that they can be burned directly into silicon, and that we can just have chips that plug into whatever it is you are building to give it these voice abilities. That way you can hopefully at least mitigate some of the more dystopian, darker possibilities.