Automated speech recognition to support the transcription of parliamentary proceedings

Use case ID: 066
Author: House of Commons of Canada
Date: 12 July 2024

Objective:

Enhance the transcription of parliamentary debates and committee meetings by implementing automated speech recognition technology, thereby producing an initial draft transcript directly from audio recordings in a timely and efficient manner.

Actors:

• Parliamentary publications staff (proceedings and verification officers, transcribers and “transeditors”)
• AI system for automated speech recognition
  Parliamentary information management system, including digital transcription and authoring modules

Prerequisites:

• Digital audio recordings of parliamentary proceedings
• Automated speech recognition technology
• Metadata to identify speakers and languages within an audio segment; alternatively, speaker diarization and language detection technologies can be used
• Integration with parliamentary digital transcription authoring systems and workflow

Scenario:

• A participant in a parliamentary plenary session or committee meeting speaks. 
• A proceedings and verification officer captures metadata to identify the person speaking, as well as the language spoken, in a parliamentary information management system’s database.
• The audio captured during the speech is recorded and ingested into the parliamentary information management system.
• The parliamentary information management system submits the relevant audio and metadata to the automated speech recognition system for processing, and the generated draft transcript is stored in the parliamentary information management system’s database.
• A transcriber/transeditor uses the digital transcription and authoring module to open a meeting segment that requires transcription.
• The transcriber/transeditor uses the relevant audio and metadata to review and edit the draft generated by the automated speech recognition system. Once this task is completed, the meeting segment is moved to the next step in the workflow.

Alternate flows:

• When automated speech recognition technology is unavailable or its integration with the parliamentary information management system is interrupted, the transcriber/transeditor uses the relevant audio and metadata to manually transcribe the audio segment before completing their task.

Expected results:

• Increased business efficiency through a reduction in the costs, resources and time needed to transcribe debates and committees for parliamentary publication
• Improved service delivery of unedited transcripts for review by MPs and faster access to official verbatim records for House and committee proceedings for MPs and the public

Potential challenges:

• AI systems may produce hallucinations that seem correct, requiring diligence on behalf of the transcriber/transeditor responsible for reviewing the generated draft transcript.
• Inaccuracies introduced by the human generation of metadata may result in small portions of content being associated with the wrong meeting segment.
• Automated speech recognition systems may be inclined to output content using US spelling, or may struggle with specific names or ridings; post-processing may therefore be needed to improve the overall quality of the draft produced.

Data requirements:

• Metadata required to integrate generated drafts into the parliamentary information management system (meeting information, person speaking, time and language spoken). Speaker diarization and language detection may supersede this requirement.