AI-powered verbatim records system (HANS)

Use case ID: 016

Author: Riigikogu (The Estonian Parliament)

Date: 5 June 2024 

Objective:

Reduce the need for human labour (stenographers) at sittings of the Riigikogu (The Estonian Parliament) and automate the working process; and assist in the preparation of committee minutes. Both outputs are reviewed by staff members (editors, committee advisers) and processed further if necessary. 

Actors:

• Members of parliament
• Editors and committee officials
• HANS verbatim records system
• Document and procedure information system
• Voting information system 

Prerequisites:

• High-quality audio recording system (part of the HANS system)
• Well-trained speech-to-text system of sufficient quality (error rate less than 10%, part of the HANS system)
• Willingness to give part of the work to the AI system 

Scenario:

1. The voice recording system records all the speeches and sends 10-minute clips to the speech-to-text system.
2. The AI speech-to-text system produces a draft transcript, identifies the speaker and sends all the prepared data to the main system.
3. For plenary sittings, the system automatically:
   1. prepares a text according to the speakers, based on the texts obtained from the different audio segments
   2. adds the necessary punctuation marks and the names of the speakers
   3. adds the agenda items according to the information received from the document management system
   4. adds the links to the votes, taken from the voting information system
   5. adds YouTube links to each agenda item and time codes of speeches.
4. Editors review the text prepared by the system and make the necessary corrections; the transcript is then automatically published in the same system on the website of the Riigikogu.
5. For committee meetings, committee officials upload a pre-recorded, high-quality audio file to the system, and the system then automatically:
   1. prepares a single text according to the speakers, based on the texts obtained from the different audio segments
   2. adds the necessary punctuation marks and the names of the speakers.
6. Committee officials review the received text, make the necessary changes and prepare the minutes of the committee. 

Alternate flows:

• The system allows for the manual production of a transcript without speech recognition. 

Expected results:

• The need for human labour is reduced. 

Potential challenges:

• The speech-to-text language model needs to be expanded with new words from time to time, and needs additional training. 

Data requirements:

• Sufficient training material is essential. 

Integrations with other systems:

• There is no direct need for integration, although any kind of integration makes the entire system more automated and further reduces the need for human labour. 

Success metrics:

• The speech-to-text error rate is less than 10%.