AI transcription for public meetings: minutes, captions, and the record

Reference article · 5 min read

Most councils, boards, and commissions record their meetings, and many stream them. Automatic speech recognition is now cheap enough to transcribe all of that audio rather than part of it, shifting clerical work from typing toward reviewing.

The technology is useful and not reliable enough to leave unattended. Errors cluster where a public body cares most: proper names, motion and vote language, and moments when two people speak at once. Before the first machine transcript is posted, an agency should be able to say in writing which document is the record.

The transcript is not the minutes

Open meeting laws in most states require a public body to keep minutes, and minutes generally describe what the body did rather than everything that was said.

• Minutes. A record of actions: items considered, motions made and seconded, votes, and decisions. In many jurisdictions they become official only when the body approves them at a later meeting.
• Transcript. A rendering of the audio. Evidence of what was said, not of what the body decided, and approved by no one unless you add an approval step.
• The recording. Audio and video are usually records in their own right, sometimes with a retention period of their own.
• Policy. Name the official record in writing, mark machine transcripts as unreviewed, and publish a way to report an error.
• Retention. Transcripts and caption files are records too. Put them on your retention schedule, log corrections instead of overwriting text silently, and confirm you can export in bulk.

Requirements vary by state and by a body's own bylaws. Confirm what applies with your clerk and counsel.

Where recognition breaks down

• Names. Member names, applicants, streets, subdivisions, and department acronyms fail most often and most visibly. Load a custom vocabulary and refresh it after elections and appointments.
• Motions and votes. Motion language is spoken fast, and roll call responses are short words that are easy to confuse. Never source a vote tally from a transcript; take it from the clerk's system or the voting hardware.
• Crosstalk. Simultaneous speech degrades both speakers, as do gavels, applause, and side conversations near a live microphone.
• Uneven accuracy across speakers. Published research has found error rates are not evenly distributed across speakers, including by accent and dialect. Public comment brings the widest range of voices into the room, so residents are often transcribed worse than officials.
• Word error rate alone is weak. One percentage counts a dropped article and a misspelled councilmember's name the same way. Ask for errors broken out on names, numbers, and motions.

Audio quality drives results more than the choice of engine. One microphone per seat, push to talk, a dedicated public comment microphone, and a chair who recognizes speakers by name will do more than switching products.

Live captions and accessibility

Title II of the Americans with Disabilities Act obligates state and local governments to communicate effectively with people with disabilities. The U.S. Department of Justice has adopted WCAG 2.1 Level AA as the accessibility standard for state and local government web content and mobile applications, with compliance dates phased by size of entity. Deadlines and exceptions change, so confirm current obligations with counsel.

WCAG 2.1 includes a Level AA success criterion for captions on live synchronized media and a separate one for prerecorded media. Expectations worth specifying:

• Latency between speech and caption, and behavior when the feed fails mid-meeting
• Captions that do not obscure the agenda, presenter, or sign language interpreter
• Marked speaker changes, punctuation, and capitalization
• Captions on public comment, not only on the dais
• Reviewed captions replacing the automatic live pass on the archived video

Automatic captions are a baseline. When someone requests an accommodation, agencies commonly provide a human realtime captioner, and an automatic feed may not satisfy the request. Machine translation of an imperfect transcript compounds its errors.

Speaker identification

Diarization separates voices and labels them generically. It does not know who anyone is. The more reliable approach maps mixer channels to seats so each microphone carries a known name; ask whether a product accepts per channel audio or a microphone to name map. Voice enrollment is possible, but storing voiceprints can raise state biometric privacy questions whose application to public agencies varies. Public commenters are a separate case: many state a name the system then gets wrong, so decide as policy whether published transcripts label commenters at all.

What to ask vendors, and how to start

• Test on your audio. Supply recordings from your own chamber, including a contentious meeting and a public comment period, not vendor samples.
• How accuracy is measured. Who produced the reference transcript, and error rates for names, numbers, and motions.
• Custom vocabulary. How many terms, who updates them, and how fast changes take effect.
• Speaker labeling. Diarization only, channel mapping, or integration with agenda and roll call systems.
• Live captioning. Whether live and archived text come from the same engine, expected latency, and failure behavior.
• Data handling. Where audio is processed, which subprocessors touch it, deletion on request, and a term barring use of agency audio and transcripts for model training without written agreement.
• Export and product accessibility. Standard caption formats and timestamped text in bulk, during the term and at termination, plus your own testing of the caption display and the staff editor.

Run the tool alongside your current process for several meetings, measure clerk time and error types, and settle what is official before publishing machine generated text.

HQT is a technology and software company serving United States state and local government agencies, with work spanning software development, systems integration, and accessibility.

Frequently asked questions