October 28, 2026
Introducing Tiro 1
Tiro 1 is the HillGenius transcription system, built only for Congress. In a public test of 100 congressional hearings drawn at random, it attributed 99.4% of words to the right speaker.
It's designed for what general-purpose transcription gets wrong in hearings: the names of members, witnesses, bills and agencies, and who said each word. It delivers a finished transcript minutes after a hearing ends, with no one editing it first.
A transcript is only useful if you can trust who said what. So we checked Tiro 1 against the official printed record of each hearing, and published every hearing, every error and the files to rerun the score.
See the test results99.4%
of words attributed to the right speaker
1,639,116 of 1,649,095 words
100 House and Senate committee hearings, scored against the official record. Notes
100 hearings, one square each
82 hearings had every speaker right. In 13 more, at least 99% of words were right.
Open any square to read that transcript as it was scored, next to the official print.
Members and witnesses
How Tiro 1 did on members of Congress, and on the witnesses who testified before them.
| Speaker | Words right | Speakers right |
|---|---|---|
| Members of Congress | 99.99%950,190 of 950,276 | 99.4%1,340 of 1,348 |
| Witnesses | 98.8%688,926 of 697,077 | 98.5%329 of 334 |
The other 8 speakers were clerks calling the roll and members answering together; all 8 were wrong. Notes
Fully automated
No one reviews a transcript before it’s delivered, and the test transcripts were scored exactly as delivered.
For hearings captured live, the finished transcript is ready a median of 11 minutes after the hearing ends, and it doesn’t slow down when many hearings run at once. Notes
Tiro 1 identified errors in the official record
In five passages, the printed record of a hearing credits the words to the wrong person. Tiro 1 named the person who was speaking.
Four turned up in transcripts Tiro 1 produced in September, and the fifth in this test. The evidence, line by line
The print says “Chairman Wicker.” Sen. Ted Budd was speaking.
Senate Armed Services · February 25, 2025
Hearing video 
The printed record, highlight added The print says “Senator Kelly.” Sen. Gary Peters was speaking.
Senate Armed Services · March 27, 2025
Hearing video 
The printed record, highlight added The print says “Mr. Fitzpatrick.” Rep. Scott Fitzgerald was speaking.
House Transportation and Infrastructure · May 14, 2025
Hearing video 
The printed record, highlight added The print says “Mr. McCormick.” Rep. Tom McClintock was speaking.
House Natural Resources · February 11, 2025
Hearing video 
The printed record, highlight added The print says “Mr. Baird.” Judge Paul Black was speaking.
House Judiciary · July 15, 2025
Hearing video 
The printed record, highlight added
How we tested it
Each step is written up in the full report, with the files to check it.
- Drawn at random
- 100 of the 143 printed House and Senate committee hearings that HillGenius had never processed, picked at random.
- Run unchanged
- Tiro 1 transcribed all 100 in production, the same system our customers use, with no one reviewing the transcripts.
- Checked against the official record
- Each hearing’s right answer came from its printed record from the Government Publishing Office, read twice independently and frozen before scoring.
- Rules set first
- The scoring rules and the scoring code were fixed before the draw.
House and Senate
Tiro 1 scored higher on House hearings. The Senate sample is smaller, at 21 hearings.
| Chamber | Hearings | Words right | Speakers right |
|---|---|---|---|
| House | 79 | 99.6% | 98.9% |
| Senate | 21 | 98.5% | 98.0% |
| All | 100 | 99.4% | 98.8% |
Where it still gets it wrong
21 of 1,690 speakers were wrong, holding 9,979 words. Most were brief, like a clerk calling the roll or members answering together.
Two causes account for most of the wrong words: witnesses we left unnamed, and two videos that started after the hearing began.
| Cause | Speakers | Words |
|---|---|---|
| Witness or guest named on the record but left unnamed | 4 | 4,489 |
| Video starts mid-hearing; a witness was given the chair's name | 2 | 3,666 |
| Clerk calling a roll-call vote | 5 | 1,661 |
| Members speaking in unison ('All') | 3 | 81 |
| Short remarks in the unidentified-speaker bucket | 5 | 61 |
| A member's brief remark, mislabeled or unnamed | 2 | 21 |
Notes
- The 99.4% counts the words that line up with the printed record: 91.2% of the 1,807,969 words in our transcripts. Most of the rest differ from the print, often because members edit their remarks before it is printed.
- Each voice’s words count for its main speaker. Counting the 3,784 words inside a voice that belong to someone else as errors gives 99.2%.
- Members and witnesses were broken out after scoring, not as part of the pre-registered test. A sitting member who testified counts as a member.
- Turnaround is for hearings HillGenius captured live from 2026-08-28 to 2026-10-27, measured from the end of the hearing to the finished transcript.
- After scoring, we corrected errors in the reference itself; none went against us. Before those corrections, the same run scored 99.1% by words and 98.5% by speakers. Both results are published.
- HillGenius ran this test. The method, the reference and every file are published so anyone can check it.
- Tiro 1 is patent pending.
See every hearing for yourself
The test results show all 100 hearings, every error and the files to rerun the score.
See the test resultsChecking our work? The full report has the method, every wrong speaker, the corrections and every file, and agents can read it as Markdown. Every file, on GitHub.
Want Tiro 1 transcripts of the hearings you follow?