Voice AI Day is baaaaack! 🎙️ 11 cities, 10 Dec
    Industry Conversation

    Jonathan Vincent Cremer, Co-Founder at Tontaube | BotCast #16

    Meet Jonathan Vincent Cremer, Co-Founder at Tontaube, who joined the BotCast #16 to answer our usual questions on building with Voice AI.

    Who are you?

    I'm Jonathan. I'm the co-founder of Tontaube. We are developing foundational text-to-speech models. I founded this startup together with my brother one and a half years ago, and we are completely bootstrapped. We just released our second model, and it's an open-weight model available on Hugging Face. We are currently in the process of talking to customers, seeing where we can improve and how it fits their needs, to sharpen the project a bit more. But we are really proud that we have now released the second model, and we go from here.

    How did you end up using voice AI?

    Initially, we wanted to build audiobooks from old texts. We started with open-weight models and APIs, but we quickly discovered that those models weren't really strong in terms of their quality, their intonation and keeping the arc of tension. So we developed our own model on the side. At first, we were skeptical that we could get to a model that was sufficiently good. But with our first big training run of $300, we already achieved quite a good result, and that led us to believe that we could actually compete in that area. Then we moved towards the voice AI space, because there's much greater potential in terms of the market at the moment, in our opinion.

    A moment of joy and pain using voice tech.

    My biggest moments of joy were when we improved our model, when we could really listen to the changes and subjectively feel a great improvement, beyond the metrics, because the metrics don't always tell the big picture, especially when it comes to voice. And because we are, of course, a bit hesitant to take our own judgment for truth, we conducted a listening study where we compared ourselves against a big competitor, and we achieved good results: we significantly outperformed them in terms of prosody, so intonation and things like that.

    The pain point was that we were a bit naive to think that if we just release a good model on Hugging Face, people will adopt it. It turns out it's much more difficult than that. That's something we really had to learn, and that was a great pain.

    What lesson would you share with other builders?

    My biggest lesson was that you shouldn't be intimidated by the big labs, who always try to convince people that what they are doing is really difficult and requires huge scale. That's true in some domains, but not in others, and you really have to differentiate between those. In audio, really small teams with small resources can compete with the big models when they really pay attention to the details, which big competitors sometimes don't do because they have legacy technology they are building on. They don't always get the first attempt right, and then they build on top of imperfect technology. That's something where a startup can really shine, I think.

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

    In the next 12 months, I think people underestimate how much of a gap there still is between a human pronouncing something and intonating it properly, and an AI system. I think there's no text-to-speech model on the market currently that really comes close to human intonation and naturalness in speech. That's more of a contrarian view I have, because many people believe that it's basically solved.

    When it comes to the five-year prediction, I think we will see a pivot towards speech-to-speech in the long term, because this guarantees really low latency and a smoother experience. You really need the audio context of the previous speech, and to make the speech that is produced depend on it, so that it really fits what the user has said. I think that's what's coming next, and something people in the space often agree about.

    TONTAUBE

    https://tontaube.ai/

    In the Voice AI Space library

    More Industry Conversations