An AI system for transcribing court hearings had to be tested on recordings from across Ireland to see whether strong accents and local turns of phrase would leave it lost for words.
The €170,000 Courts Service artificial intelligence project was trialled using audio from courthouses nationwide amid concerns that accents, dialects and “localisms” might prove confusing.
However, testing found the AI system performed strongly overall and made just 585 “word errors” across more than 30,000 words that were manually checked.
The Courts Service said this was a “word error rate” of 2.2 percent, with poor audio proving a far bigger impediment than regional accents.
The system, which is called the DAR AI Assistant, is now being moved to a small-scale live pilot for use by judges and Courts Service staff.
The AI is designed to make lengthy court recordings easier to search, allowing people to find testimony or sections of hearings without having to listen back through a full set of proceedings.
A spokesman explained: “For example, reviewing a five-hour court sitting to identify a specific exchange or testimony can be time-consuming.
“The DAR AI Assistant helps users find and access those sections more quickly.”
The spokesman said the system was not intended to replace external transcription services but instead to become a useful reference tool.
It will only be used by Courts Service staff and judges, who will be trained on both the capabilities and limitations of AI-generated transcripts.
The spokesman said: “Human review remains central to the process and users will be made aware that AI-generated outputs should not be relied upon without validation against the original audio.”
Internal records said the project would face additional challenges, aside from accents including courtroom sound quality.
A business case said personal information in the transcripts would be carefully stored and processed and that a data protection impact assessment had taken place.
On project disadvantages, it said: “Introduction of the [system] will require judges and staff to familiarise themselves with new functionality and ways of working, which may initially slow adoption.”
A list of risks said data quality would be closely monitored as “inaccurate transcripts or delays in audio availability could reduce confidence in the solution.”
Another risk assessment document said a separate sound modernisation project was underway to improve audio quality. To deal with “poor quality audio, Irish accents, and localisms,” it said the AI could be provided with bespoke dictionaries and tweaked over time.