In a busy healthcare environment, staff were still relying on manual note-taking. It was slow, tiring, and easy to get wrong. The obvious solution seemed to be voice recognition — but there was a catch. The recordings were in a less common Scandinavian language and used an old, heavily compressed audio format. Most off-the-shelf speech-to-text tools simply couldn’t understand it.
That’s where Zfort stepped in.
Instead of trying to force a generic solution, the team decided to build a custom voice recognition system from the ground up. They started by recording native speakers reading carefully selected texts. This helped create a dedicated speech dataset that matched the client’s real audio conditions.
Training the AI wasn’t easy. The audio quality was poor, the vocabulary included specialized medical terms, and there was very little public data available for this language. Through repeated training, testing, and fine-tuning, the model slowly learned to recognize speech that other systems ignored.
Eventually, it worked. Spoken notes could now be automatically converted into accurate text, even from compressed and noisy recordings.
For the healthcare provider, this meant faster documentation, fewer errors, and less administrative burden for staff. Instead of typing or dictating to another person, they could speak naturally and let the AI handle the rest.
This project showed that when real-world conditions don’t fit standard tools, a custom AI approach can make all the difference.
Never miss a story from us, get weekly updates in your inbox.