A full stop every 30 seconds
locl cuts short audio clips from South African radio: a caller's question, a minister's answer, the line that made a host laugh. Each clip is at most three minutes long, cut from the 30-second windows we transcribe. A clip only works if it starts where a thought starts and stops where it ends.
In September about a third of our clips stopped in the middle of a sentence. The cause was a full stop that isn't in the broadcast.
How a clip is cut
A model reads the last half hour of transcript and picks a moment. It returns the windows the moment spans, plus a short phrase copied from where it starts and one from where it ends. Our code finds those phrases in the transcript's word timings, cuts the audio there, and then moves each cut to the nearest sentence boundary so it doesn't land inside a word.
Every step after the model's pick trusts the transcript's punctuation to say where a sentence ends.
The seam
Each station's audio reaches the speech model (NVIDIA's Parakeet) as a 30-second window, and each window is transcribed on its own. The model has no idea that the window before it, or the one after, exists. So it writes every window as if it were a complete passage.
We measured that on 3,000 windows from yesterday:
| What the transcript says | Share of windows |
|---|---|
| The last word ends in a full stop, question mark or exclamation mark | 80% |
| The last word ends at 30.0 seconds, so speech was still going when the window closed | 84% |
| The first word starts with a capital letter | 99% |
We first measured this in September and got the same three numbers. A presenter does not finish a sentence every 30 seconds on the dot. The window closes mid-word, and the model puts a full stop there anyway, then opens the next window with a capital as if a new sentence had begun.
Three things believed it
Three parts of the clip pipeline took that punctuation at face value.
The picker's instructions said to leave the end phrase empty if the excerpt "ends at a full stop", since the clip would then end cleanly. Nearly every excerpt did.
The end-of-clip snap looked back for the last full stop and found one on the window's final word. It left the cut where it was, on the window edge.
The start-of-clip snap could never see that a window opened mid-sentence, because the window's first word always had a capital.
Over a week of feed clips, 35% of the clips that spanned several windows and 43% of the single-window ones ended on the raw window edge. About five times in six that edge was in the middle of speech.
The padding
There was a second, smaller fault. When a clip was cut at a quoted phrase, the code added 300 milliseconds before it and 700 after, so the first and last words weren't clipped. Radio speech is fast and those pads were fixed. In 51% of clips the 300 milliseconds reached back into the end of the previous word, and in 62% of the clips trimmed at the end the 700 milliseconds caught the start of the next sentence.
The first fix
The first round of changes went out on 3 September:
- A full stop on a word that ends at the window edge is not a sentence end.
- Whether a window opens a new sentence is decided by the previous window's last word, not by this window's capital letter.
- The pads stop 40 milliseconds short of the neighbouring word, whatever the fixed amount would have been.
- Every clip fades in over 20 milliseconds and out over 60, so a cut between two words never clicks.
Twelve hours later, against the week before:
| Before | After | |
|---|---|---|
| Clips opening on the tail of the previous word | 48% | 0% |
| Clips ending on the start of the next word | 17% | 0% |
| Clips ending on a real full stop | 19% | 77% |
| Clips ending on a raw window edge | 34% | 22% |
Crossing the seam
The 22% still ending on the window edge had one cause. In 73% of them the picker had copied the window's last sentence as its end phrase, fake full stop included, and a matched phrase was treated as final.
So the next change, a day later, lets a cut leave its window. When a clip ends on a raw window edge, the code reads the next window and runs forward to the first real full stop, up to 10 seconds. We picked 10 from the data: the next real sentence end sat within 6 seconds of the edge 48% of the time, within 10 seconds 71% of the time, and within 12 seconds 86% of the time. A clip that starts on a raw edge runs back into the previous window, up to 6 seconds, to the first word of the sentence in progress.
Neither crossing goes into an advert, and neither takes a clip past three minutes.
We also started recording, on every clip, where each edge came from: the picker's phrase, the window edge, a snap within the window, or a snap across the seam. Our server logs keep about an hour, and we wanted to review the change a day later.
What we had wrong about starts
Two rounds of fixes had cleaned up the ends. The starts had not moved: 27% of clips still opened mid-sentence.
In mid-September we wrote a read-only audit. It rebuilds the words inside each of the last 1,000 clips from the word timings and judges each edge from timing and punctuation. An edge resting on a window's last word is marked unknown, because that punctuation can't be trusted.
It showed two things. A quarter of the mid-sentence starts followed a comma (the bar marks where the clip began) ("So, Lester, | one of the biggest issues…") and were deliberate. And almost every genuinely broken start had come from the picker's own start phrase, which began a word or two into a sentence ("I've had a few | encounters like that at Home Affairs"). The snaps stood down for a matched phrase, so nothing we had fixed so far could reach those starts.
The audit had a bug of its own. It placed each word a few hundred milliseconds late, so the previous word was dragged into the clip and read as a broken start. We found it with a check that needed no ground truth: a start produced by a snap lands on a sentence start by construction, so none of them should read as broken. With the bug, 28 did. After the fix, none did, and the numbers moved:
| First count | Corrected | |
|---|---|---|
| Clips starting mid-sentence | 27.4% | 22.1% |
| Genuinely broken starts, with no comma before them | 20.6% | 13.8% |
| Clips ending mid-sentence | 14.2% | 2.9% |
The fix for starts walks a picked start phrase back to the first word of its sentence, within the same window and at most 6 seconds. It leaves a start after a comma, colon or semicolon alone, and it won't move a start just to put back a leading "And", "So" or "But".
Where it stands
We ran the audit again this morning on the 1,000 most recent clips. For about 15% of edges at each end the punctuation can't settle it, because the edge sits on a window boundary. Of the rest:
| Share of edges | |
|---|---|
| Starts mid-sentence | 4.5% |
| Genuinely broken starts | 1.7% |
| Ends mid-sentence | 5.2% |
The broken starts fell further than the walk-back alone predicted, from 13.8% to 1.7%. We also moved the picker to a different model in October, so we can't say how much each change contributed. Ends went from 2.9% to 5.2%, and we haven't looked into why yet.
If you chunk audio for a speech model
- Look at the word timings at the chunk edge before trusting the punctuation there. A model that sees one chunk will write a complete passage, whether or not the speaker finished.
- Anything downstream that reads sentence boundaries inherits the fault, including a language model you hand the transcript to.
- Measure where the bad edges come from before fixing them. Our first two rounds fixed real problems, but not the one behind broken starts.
- Store how each output was produced next to the output. Logs expire before the review.
- Give an audit a check it must pass by construction. Ours caught its own alignment bug that way.
You can hear the clips on the feed.