AI could soon transcribe court, parliament sittings in Maltese

The government is seeking to develop a centralised Maltese speech-to-text engine that would automatically transcribe court sittings, parliamentary proceedings and, eventually, audio uploaded by the public.

A tender published this month calls for the development of an AI-powered transcription tool capable of handling Maltese, English and the kind of Maltese English code-switching common in everyday local speech.

The project, valued at an estimated €147,750 excluding VAT, forms part of the Recovery and Resilience Plan measure for the digitalisation of the justice system and may be partly financed using EU funds.

Court transcriptions are currently carried out manually by transcribers using audio recordings of sittings – a process described as “costly and time-consuming”. The new system is intended to automate and optimise that work, while also extending similar capabilities to parliament and a public-facing transcription portal.

The tender says the traditional transcription role in the courts would shift from manual typing to reviewing and correcting machine-generated transcripts through a dedicated ‘smart editor’.

That editor would allow reviewers to listen to specific parts of the audio, correct the corresponding text, adjust speaker labels and feed corrections back into the system.

The system is expected to distinguish between different speakers, including when they share a single microphone. However, it will not be required to automatically identify people by name: generic labels such as Speaker 1 and Speaker 2 would be enough, with reviewers able to manually rename speakers where necessary.

A focus on accuracy

One of the most ambitious requirements is accuracy. The system must reach 85 per cent accuracy in its first year, 90 per cent in its second year, 95 per cent in its third year and more than 95 per cent in years four and five. Accuracy will be measured through word error rate across courts, parliament and other use cases.

The tender specifically requires the engine to support Maltese as its primary language, English, and mixed Maltese English speech within the same recording, including mid-sentence switches. It says the system must be able to deal with courtroom recordings, parliamentary audio and other environments involving multiple speakers and background noise.

A public portal is also envisaged, allowing users to either upload audio files for transcription or use a microphone for real-time transcription. The resulting text would be provided directly through the portal, by email or as a downloadable document.

System must comply with GDPR

The tender requires safeguards to prevent abuse and control costs, including limits on audio length, file sizes and rate-limiting. It also states that audio uploaded through the public portal must not be stored once processing is complete.

For courts and parliament, the system must include audit trails and access logs sufficient to support legal audit and evidentiary requirements.

The system must comply with GDPR and national data protection law.

The successful contractor will have eight months to develop, deploy, install, test and commission the software, followed by five years of maintenance and support.

The contract will be awarded solely on price to the cheapest offer that satisfies the administrative and technical criteria. Bidders must show experience in natural language processing or large language model projects and must hold ISO 27001 information security certification or an equivalent recognised standard.

The deadline for submissions is June 8. A clarification meeting will be held online on May 20.