What's new
v1.3.4 · 25 September 2026Every stream was transcribed again with NVIDIA Parakeet in place of the older Whisper. This page measures what that changed across 778 streams, and lists the other improvements shipped alongside it. Figures were measured on 2026-09-25.
Lines that repeat verbatim. Whisper padded silence and music with the same phrase over and over.
Identical lines in a row. The worst went from 7,086 to 15.
Out of 778.
657,840 filler lines removed.
Fewer duplicate lines than before. 21 are slightly higher and 7 are level.
Fewer invented lines, everywhere
Whisper's failure mode on long streams is to loop: the same short phrase written down hundreds of times over silence, music or game audio. The measures used are the share of lines that repeat an earlier line exactly, and the longest unbroken run of identical lines.
How many streams sit at each duplicate level
Streams in each band of duplicate lines. Bars moving left is the whole story.
View as table
| Band | Whisper | Parakeet |
|---|---|---|
| 0-5% | 20 | 193 |
| 5-10% | 39 | 440 |
| 10-20% | 288 | 132 |
| 20-35% | 237 | 9 |
| 35-50% | 104 | 2 |
| 50-75% | 65 | 2 |
| 75-100% | 25 | 0 |
How long the worst loop got
Streams by their longest run of identical lines. Loops of 500+ lines went from 77 streams to 0.
View as table
| Longest run | Whisper | Parakeet |
|---|---|---|
| <5 | 25 | 557 |
| 5-19 | 98 | 221 |
| 20-99 | 352 | 0 |
| 100-499 | 226 | 0 |
| 500-1.9k | 58 | 0 |
| 2k+ | 19 | 0 |
Better in every year of the archive
Average duplicate lines per stream, by year.
View as table
| Year | Streams | Whisper % | Parakeet % |
|---|---|---|---|
| 2018 | 4 | 24.4 | 7.9 |
| 2019 | 224 | 30.8 | 9.8 |
| 2020 | 167 | 27.0 | 5.7 |
| 2021 | 42 | 19.7 | 6.1 |
| 2022 | 49 | 30.0 | 6.0 |
| 2023 | 28 | 26.6 | 6.4 |
| 2024 | 65 | 29.4 | 6.2 |
| 2025 | 154 | 25.3 | 8.2 |
| 2026 | 45 | 21.0 | 7.2 |
The eight worst loops, fixed
Longest run of identical lines, before against after.
- 2025-10-17 - DRIVETIME fastBLAST7,086 → 5 in a rowduplicate lines 73% → 5%
- 2025-08-28 (2) - MGS3 Remake6,922 → 3 in a rowduplicate lines 72% → 8%
- 2024-12-14 - Trash Division4,414 → 2 in a rowduplicate lines 83% → 6%
- 2022-08-08 - TOTAL TRANSPORT 44,027 → 12 in a rowduplicate lines 83% → 8%
- 2024-10-11 - Disturbing and Unacceptable Scenes4,027 → 2 in a rowduplicate lines 86% → 4%
- 2025-09-26 - MGS44,010 → 3 in a rowduplicate lines 82% → 7%
- 2025-03-25 - Fallout_ Lou Begas - No Damage Speedrun any% - DON'T EVER TRY THIS, YOU WILL FAIL3,761 → 8 in a rowduplicate lines 60% → 7%
- 2024-10-14 - Disturbing and Unacceptable Scenes3,369 → 3 in a rowduplicate lines 56% → 7%
What the old model was making up
2026-09-15 (2) - More misery
One flight-simulator radio phrase and the number after it were written 1,151 times each, in strict alternation, between minutes 120 and 160. Parakeet has ordinary speech there instead.
2025-04-15 (2) - Verdansk '25
The old transcript ran to 284 minutes. The video is 3 hours 45 minutes, so the last 59 minutes of that transcript described nothing that exists.
2021-08-13 - Shlug ~ Life
"Thank you." written 149 times over roughly an hour of silence.
2023-06-04 (1) - Wash and Wax
Ended with "Transcription by CastingWords" repeated 30 times, a well-known Whisper artifact.
Word count fell 11% overall (15.1M to 13.5M). That is mostly the invented filler being removed, so a lower count is the intended result. The measures count lines, not correctness, and a real recurring bit (a soundboard clip, a repeated joke) can legitimately repeat, which is why the after figure is not zero.
Shipped alongside
Copy an answer to share
A Copy answer button on the Ask page copies the answer as italic markdown, with every citation linked to the exact moment on YouTube. The links are written so Discord does not unfold each one into a video player and fill the channel.
A summary for every stream
Open a stream and its summary is there, or you can ask for one to be written. Nearly every stream now has one, and new streams get theirs automatically as they are added.
A rebuilt index
The whole archive was re-chunked and re-embedded from the new transcripts, so search and Ask work from cleaner text: 785 streams and 59,280 passages. Before, there were 780 streams and 53,471 passages, a fifth of them padding.
Characters, relearned
The four character profiles were rewritten from the cleaner transcripts, so they describe how each one actually talks instead of picking up the old model's mistakes.
Made for phones
Library rows stack on a small screen, the heatmap swipes and its days are big enough to tap, and typing in a filter no longer makes iOS zoom in. The nav fits five buttons on a phone.
A sharper library
The heatmap can be saved as an image that follows the active host, tag and date filters, drawn in the same style as the app. Hide empty condenses it to only the months that had a stream.
See where a stream came from
Each stream page shows which speech-to-text model made its transcript and which AI wrote its summary, so it is clear what you are reading.
Better tags
Episodes of Computer Chronicles and Net Café that were missing their tag can now be found from the words around them, so the archive's tags are more complete.
Testing before changing the AI
A newer AI model is tried on the same streams and the same questions as the current one before anything changes. The latest test found it more careful about not inventing an answer and better at covering a long stream, but about twice as slow.
Limits worth knowing about
Answers can be wrong
Answers come from an AI reading transcripts, and the transcripts are machine-made too. Open the linked moment to check a quote before relying on it.
Links start a little early
A citation link opens the video at the start of the passage it cites, about 5 seconds early for context. Passages run a couple of minutes, so the line being quoted can come a little way in.
A few streams are being redone
A handful of transcripts are being re-transcribed, so they may still read like the older version until they are replaced.
Not every stream has a video link
Some streams have no YouTube link on file. Their citations show the stream and time, but cannot jump to a video.
A few streams have no summary
Three are too short to summarise: a 15-second clip, one with no speech, and one under 200 words.