Speech recognition is good at words and bad at names. It has no way of knowing that the sound it just heard was a person, let alone which person, so it writes down something plausible and moves on. Across 460 episodes that adds up, and this week the archive went through 2,146 name corrections across 341 episodes.
The worst offender was our own. Lyndell Murray's name had been going in as Lyndall — 876 times, across 205 episodes, against fifty that had it right. If a host's name can go astray that consistently, nobody else stood much of a chance.
Not everything turned out to be a spelling mistake. A few times two people standing together had been quietly fused into one person who never existed. Once or twice a parkrun had been mistaken for a parkrunner. Those are the ones that are oddly satisfying to put right, because somebody gets their own line back.
Some things were left exactly as they were said, on purpose. If the show calls you by a short name, that is what the Log says — the point is to be faithful to the programme, not to file everybody under their formal spelling. A record of a conversation should sound like the conversation.
None of this touches the audio. The episodes are untouched; what changes is the written record, and whether searching for your name actually finds you — which, if you have ever gone looking and come up empty, may be the whole point of the exercise.
If your name is still wrong, the feedback form remains the fastest way to fix it. A name reported there is corrected at the source and carried into every episode it appears in, past and future.
Episode 457 now carries the Verified badge. Someone sat down with the audio and the transcript and went through it — every name, every spelling, every segment heading, against what is actually said. That took hours, and it is the only thing the badge has ever meant.
What the check covers, and what it does not. The work goes into the things that carry meaning: who is speaking, which parkrun they are reporting from, how a name is spelled, where one part of the episode ends and the next begins. It is not a word-by-word audit against the audio. Speech recognition tidies as it goes — a false start vanishes, an "um" is dropped, a sentence someone abandoned halfway is written as though they finished it — and a Verified transcript still carries those. Read it as a faithful record of what was said and who said it, rather than as a court transcript.
The earlier note on the three labels said that no episode was Verified. That was true when it was written. It is now true of 456 of them.
No Verified badge means no human has checked it. An episode showing only Enhanced transcript has been through the automated editing pass and nothing else. The badge is added only when the work has actually been done, so its absence is not an oversight or a backlog marker — it is the honest state of that episode.
Verifying an episode is slow by nature — experience says about twice as long as listening to it, and then some. So the bulk of the effort will go to new episodes, checked as they are released. The back catalogue is not being written off — older episodes will be verified too, as time and resources allow, and every one that is will gain the badge. But there are 456 of them, so expect the number to move slowly and expect Enhanced to remain the normal state for a long while yet.
In the meantime the feedback form is much the faster way to get something fixed — a name you report is corrected at the source and carried into every future episode, whether or not that episode has been verified.
Using the Log with a screen reader, a magnifier or a keyboard
The point of the Log is that the show can be read. That is no use if the pages are hard to get around, so the site is built to stay out of your way.
Every page is laid out in a sensible order with a proper heading structure, so you can jump through it rather than wade. There is a skip link to the content at the top of each page.
Photographs are described by what is actually in them — "Ali and Lynne at Oaklands College parkrun", not "artwork for episode 457". Where an episode has no photo of its own, it says so instead of leaving you wondering.
Timestamps read as what they do: "Listen from 18 minutes 2 seconds", rather than a bare number.
Text is readable against its background. The show's amber was too pale for body text on our light background, so links use a darker shade of it. The look did not change; the readability did.
Light, dark or whatever your system prefers, remembered between visits — and no animation at all if your system asks for that.
Everything works from the keyboard, with a clear outline showing where you are, and an ordinary audio player your browser already knows how to drive.
Enlarging the text reflows the page instead of pushing it sideways.
A transcript that only gives you the words leaves a lot of the show behind. These are written so you can follow an episode without hearing it.
You are told when people laugh. And when there is music, applause, or everyone talking over each other at once — marked where it happens, not summarised at the end. There are 3,344 of these moments noted across 404 episodes.
Nothing is quietly missing. Sometimes the transcription simply fails on a stretch of audio, usually when someone is speaking another language. Rather than close the gap and hope you do not notice, the transcript says something was said here and gives the length. It happens 346 times, about 92 minutes in total.
Other languages are translated, with the original kept beside it, so you get the words and the fact that they were not in English.
Press any timestamp and the audio jumps there, if you use both together.
Every episode says how far it has been checked rather than leaving you to guess.
Every episode carries a label saying how far its transcript has been taken. The three mean quite different things, and the difference is worth knowing before you rely on one.
Transcript — the words as speech recognition heard them. Accurate more often than not, but speakers are numbered rather than named and nothing has been looked at.
Enhanced transcript — an automated editing pass has named the speakers, marked the segments, described laughter, music and crosstalk, and corrected misheard names against the roster and the parkrun event list. A large improvement. No person has read it.
Verified — someone has listened to the audio and confirmed the names, the spellings and the segment headings. This is the only label that means a human checked.
Right now every episode is Enhanced and none is Verified. We would rather say that plainly than let a label imply a check nobody has done.
So please expect errors. Speech recognition mishears names constantly — a parkrun becomes a place that does not exist, two people talking over each other become one, an unfamiliar surname becomes three different spellings in the same episode. The editing pass fixes a great many of these and cannot fix them all, and it occasionally introduces one of its own by being confident about the wrong thing.
If you spot something wrong, please tell us. A misheard name, a speaker attributed to the wrong person, a roving report filed under the wrong parkrun, anything that reads as nonsense. Corrections are applied at the source and carried into future episodes, so a name you fix once stops being wrong everywhere — including in episodes not yet edited. That is genuinely how this gets better.
The last of the back catalogue has been through the editing pass. All 457 episodes — 456 released plus the "Episode 59:59" bonus — now have an enhanced transcript with speakers named, segments marked and laughter, music and crosstalk described. No automatic transcripts remain.
Episode 113 has none, and never will: it was never released.
What that changes for a reader: a search now reaches every word the show has broadcast since January 2016, and the people and events indexes describe the whole run rather than the parts that happened to be processed.
Event pages for parkruns in South Africa, Namibia and Eswatini now carry the event's own Facebook page, and its own X account where it has one — 285 links across 227 events.
Only accounts belonging to the event itself are shown.
The rest of the world follows once the links are available.
Legs Eleven, with roving reports from Mandeville, Richmond Olympic, Irchester Country and Spanish Banks Beach — online within a few hours of release, with a full transcript.
Ten years of the parkrun Adventurers podcast, written down and made searchable.
The show has been going since January 2016 and there are hundreds of hours of it. All of that sits inside audio files: who reported from which parkrun, who was interviewed and about what, when a segment first appeared, the name of the volunteer at an event on the other side of the world. None of it could be looked up. You either remembered the episode or you did not.
So this is what the Log is for:
Read an episode instead of listening to it. Every released episode has a transcript with the speakers named and the segments marked. For anyone deaf or hard of hearing that is the difference between the show being available and not.
Find a moment again. Search runs across every transcript, person, event and segment at once. Press a timestamp and the audio jumps there.
Follow a person or a parkrun. Each roving reporter has a page listing everywhere they have reported from; each event has a page listing every report filed from it.
It is beta, and the label is meant literally. The back catalogue is still being worked through, so some episodes have a full transcript and others still have the raw automatic one — each page says which. Names will be wrong in places. Search will occasionally surprise you. Every episode page tells you honestly how much has been checked rather than presenting everything as finished.
Corrections are the most useful thing you can send. They are applied at the source, so a name fixed once stops being wrong everywhere, including in episodes not yet processed. Tell us what is wrong.
An independent project by listeners, not affiliated with or endorsed by parkrun.
