Jeenius says AI still lacks the data to interpret real-world speech

Healthcare interpreting exposes AI’s language gap

Jeenius tested leading AI models while developing its AI interpreter for healthcare use. The company says the results show a gap between what the public expects AI to do and its readiness for real-world interactions.

The announcement points to interpreting as one example of that gap. It asks why Apple’s AirPods offer interpreting in nine languages when AI systems are often described as knowing hundreds of languages. Jeenius says only 20 languages met its accuracy threshold in its healthcare speech-to-speech tests.

According to Baker, healthcare conversations require more than converting words from one language into another. In a busy emergency department, an AI system would need to distinguish between speakers, understand intention, account for cultural nuance and respond naturally while television and other conversations create background noise.

Jeenius identifies a shortage of conversational data

Baker argues that the central problem is data. Text can train AI to perform tasks, but natural interaction requires audio and video showing people speaking and responding in authentic settings.

“Text trains AI to do tasks,” Baker said, according to the Jeenius announcement. “To engage with humans, AI must be trained on authentic audio and video data where people are speaking and responding naturally.”

The company says frontier models have already drawn heavily from the same public internet. It cites an estimate that 90% of AI training data is English and comes largely from North America and Europe, while 2% comes from the African continent. Baker says the data AI needs must be created rather than simply found online.

A multilingual data network for AI developers

Jeenius says it has spent years building a community of tens of thousands of qualified linguists across hundreds of languages. Its Experts-in-the-Loop network creates authentic conversational data, builds domain-specific datasets and benchmarks model accuracy for AI labs and companies.

The company says this work is intended to provide data that better reflects how people communicate across languages and settings. Jeenius describes its offering as multilingual data infrastructure for developers of large language models and small language models.

For users, the company’s findings suggest that broader language coverage depends on the quality and range of training data, especially in high-stakes settings such as healthcare. Baker summarized the company’s position by saying, “The winning AI won't be the biggest model. It will have the most reach: engaging a patient in Guatemala, a customer in Lagos, a family in Kyiv.”