How We Accidentally Started Building Pronunciation Intelligence

One of the more interesting projects I’ve worked on didn’t begin as a startup idea. It didn’t begin as research. It began while we were building conversational AI with the Chicago Bears Digital Team.

At the time, our focus wasn’t pronunciation. We were building conversational experiences for fans. Like most AI teams, we spent our time thinking about conversations, latency, user experience, and making interactions feel more natural.

The bug that wasn’t just a bug

Then we started testing with real player names. Everything worked remarkably well until the AI introduced an athlete. It consistently mispronounced names that weren’t common in English. Sometimes it was close. Sometimes it wasn’t even recognizable.

At first we treated this like any other software bug. Look up the pronunciation. Update the phonemes. Test again. Move on. After doing that enough times, we realized we weren’t fixing names anymore. We were repeating the same workflow over and over again.

Every correction required someone to verify how that athlete actually pronounced their own name. Then we had to understand the phonetics. Then generate phoneme mappings. Then test them. Then test them again.

Eventually someone asked the obvious question.

Why is everyone solving this individually?

Developers were searching YouTube. Broadcasters were listening to interviews. Journalists were checking Google. Everyone had invented their own process because there wasn’t a trusted source of truth.

From corrections to infrastructure

That was the moment the project changed. Instead of correcting pronunciations one player at a time, we started thinking about infrastructure. What if there was a verified pronunciation database? Not just names—verified audio, structured phoneme mappings, developer-friendly APIs. One source of truth that everyone could build upon.

The more we researched, the more obvious the problem became. Sports are global. Athletes come from every language, culture, and country. Most speech systems perform reasonably well with common English names. They’re much less reliable when names fall outside those patterns.

Correct pronunciation isn’t simply a speech problem. It’s an identity problem.

Names matter.

Finding the names speech systems miss

To support the research, we partnered with Sportradar. Their data gave us access to millions of athlete names across more than 500 leagues dating back to the late 1800s. That allowed us to identify uncommon and diverse names at scale instead of manually hunting for them one at a time.

From there we began pairing verified pronunciations with structured phoneme mappings and reference audio. The long-term vision became surprisingly simple. Instead of every developer solving pronunciation independently, we wanted infrastructure that anyone could use.

What a phoneme mapping looks like

Most people only see a name. Speech systems need something much more structured. A single athlete’s name moves through several layers before a voice interface can say it correctly.

<phoneme alphabet="ipa" ph="dʒoʊl">
  Joel
</phoneme>

<phoneme alphabet="ipa" ph="i-jeɪ-bu-ni-weɪ">
  Iyiegbuniwe
</phoneme>

One request. One answer. No guessing. That’s the shape of the infrastructure we wanted to build.

Phonemes as a Service

The project eventually evolved into what we internally called “Phonemes as a Service.” An API capable of providing verified pronunciations and phoneme combinations to developers building voice interfaces, sports products, conversational AI, and media experiences.

The goal wasn’t simply better speech synthesis. It was making AI more respectful, more inclusive, and more accurate—so that the next team building a voice product wouldn’t have to start from zero.

What stayed with me

What stayed with me wasn’t only the phoneme work. It was the realization that many meaningful improvements in AI do not come from larger models. They come from solving the small problems everyone quietly accepts.

Names are one of those problems. People notice when software gets them wrong, and rarely notice when it gets them right. That is exactly why I still find pronunciation intelligence worth exploring.

Most AI systems guess. I’m interested in helping them know.