Beyond the algorithms and models – human voice in AI solutions

Most conversations about artificial intelligence (AI) begin with algorithms, models, and computing power. But in reality, every successful AI solution starts somewhere much more human — with people willing to share their voice.
Anyone who was ever amazed by the mumbling of Lithuanian streets names while using GPS, was unable to dictate and send SMS in Lithuanian while driving, voice add a calendar event in their native language will appreciate the need of linguistically diverse AI solutions.
Recently, we had the opportunity to contribute to one of the most ambitious Lithuanian language technology initiatives to date: the creation of a large-scale Lithuanian voice library for AI and speech recognition systems. The project aimed to build a publicly available Lithuanian voice library containing up to 10,000 hours of recordings — the largest in the country’s history.
Our role focused on coordinating and delivering a “read speech” collection effort: gathering over 5,000 hours of spoken Lithuanian from people across different regions, age groups, genders, and speaking styles.
At first glance, this may sound straightforward: “Ask people to read texts aloud.”
In practice, it turned out to be one of the most fascinating large-scale service designs and engagement challenges we have worked on.

AI does not understand Lithuanian
This project reminded us of an important truth that is often overlooked in AI discussions – AI does not learn languages. It learns patterns. And patterns only become meaningful when the data reflects real life.
Speech pace changes with age. Pronunciation shifts between regions. Background environments vary. Reading styles differ. If an AI system only hears a narrow slice of society, it will only work well for a narrow slice of society.
Creating a truly useful voice library therefore became not only a technical mission, but also a societal one – how do we ensure that digital Lithuanian includes everyone?

5,000 hours means much more than 5,000 hours
Projects of this scale are usually described through numbers: hours collected, participants onboarded, recordings validated.
But numbers alone hide the operational reality behind them.
To make the initiative successful, we needed to think simultaneously about:

  • participant motivation,
  • accessibility,
  • communication clarity,
  • trust,
  • technical simplicity,
  • demographic diversity,
  • quality assurance,
  • and long-term engagement.
Collecting speech data is not just a data problem. It is a human participation problem. The success of the project depended on whether people from all parts of Lithuania would actually want to contribute their voices.
And they did. Students, seniors, professionals, people from cities and smaller towns — over 8,000 people contributed to building a digital foundation for the future of Lithuanian language technologies.Designing participation, not only processes
One of the most interesting discoveries during the project was how strongly participation depends on friction.
Even highly motivated people disengage when:
  • instructions feel unclear,
  • the process takes too long,
  • technology becomes confusing,
  • or the purpose feels distant and abstract.
This meant the project required much more than operational coordination. It demanded continuous thinking about user experience.
How do you explain a complex AI infrastructure initiative in a way that feels personal?
How do you help someone understand that their voice recording today may improve public digital services tomorrow? How do you convince someone, especially a senior citizen, that their voice will not be “stolen” or used for fraudulent purposes?
In many ways, the project resembled the design of a public digital service more than a traditional data collection exercise.
AI may be built on data. But meaningful AI is ultimately built on people.
And sometimes, progress starts with something as simple — and as powerful — as asking people to speak.

RELATED STORIES