Meet Laurent Mazare, co-CTO at Gradium.AI, who joined the BotCast #3 to answer our usual questions on building with Voice AI.
Who are you?
My name is Laurent. I'm one of the cofounders of Gradium. At Gradium, we are a foundation model company. We build every kind of voice AI models. That means text-to-speech, speech-to-text, and even speech-to-speech models nowadays.
How did you end up using voice AI?
Back in 2023, we started an open science research lab called Kyutai, and that's where I met Neil, my cofounder. One of our biggest projects there was Moshi, which was the first speech-to-speech, open-source model that we published in 2024, and that's what got me started in the field.
A moment of joy and pain using voice tech.
As we are a foundation model company, we train our own models, and that's very much related to that. You get lots of joy when you see your model learning and getting new capacity based on different languages and all that. And the moments of pain are also related to that. From time to time, you see your model not learning, the curve doesn't go down, and you don't really understand why. At that point, you just have to look at the code and debug stuff until it works.
What lesson would you share with other builders?
I would say that observability is key, and that applies both to technical and nontechnical things. You have to measure things properly so that you can take enlightened decisions. For example, if you have a product that relies a lot on latency, you want to check that latency is good and have proper numbers around it. And that also applies to nontechnical where you have to measure what makes your success and optimize for that.
Where do you see the industry in 12 months and in 5 years?
Over the next 12 months, I think we will see the advance of the first speech-to-speech models. They won't take the place of the cascaded solution for every use case, but they will start to be quite appealing. Over the next 5 years, I think we'll see voice everywhere, meaning that the way you interact with any form of AI agent will be using voice.
