What 693 Practice Recordings Tell Us About How People Speak

By Noam Shemla · Measurements

TL;DR Across 693 recordings, the median pace was 129 words a minute, the median filler rate was 1.9 percent of words, and "so" accounted for about half of all the fillers we detected.

Every number on this page comes from recordings people made on Exprea to practise: a talk of one to three minutes, recorded on their own device, then measured. We are publishing the aggregate because nobody else has this kind of data about ordinary people practising, and because it is more useful than another article repeating the same advice.

What we looked at

The set is every practice recording made between 12 June and 25 September 2026 that was at least 30 seconds long, leaving out recordings made by our own staff. That is 693 recordings from a few dozen speakers. The median recording was 101 seconds long. 143 of the recordings were turns in a live speaker battle, where two people take turns on a topic; the rest were solo practice.

For the speech measures below we used only the 614 recordings that went through our measured path, which timestamps every word. Almost all of them were in English.

• 693 recordings. June to September 2026, at least 30 seconds each, staff excluded.
• dozens speakers. A small group. Read every number here with that in mind.
• 129 median words a minute. Middle half: 116 to 140.
• 1.9% median filler rate. About one filler every 50 words.
• 7 pauses a minute. Median count of silences longer than half a second.
• 101 s median length. Most recordings are short practice talks.

The headline numbers from the full set.

People speak more slowly than the advice says

The figure most often repeated about speaking pace is around 150 words a minute. Our median was 129, and the middle half of recordings sat between 116 and 140.

That also puts half of our own speakers below the band where Exprea's pace score gives full marks, which is 130 to 160. We have not changed the band because of this. A practice recording is a person alone, often speaking a talk for the first time, and people slow down when they are thinking. But it is a good reason to read a pace score as a prompt rather than a verdict.

"So" is the filler, not "um"

We counted every filler our instrument detected: 2,445 in total. The ten most common were:

"Um" is the word everyone is taught to avoid, and it is close to the bottom of the list. The words that fill the gaps in these recordings are ordinary words: "so" at the start of a sentence, "like" in the middle of one.

One important limit. Our instrument counts every "so" and every "like" as a possible filler, including the times they carry real meaning, as in "so that it works" or "I would like to". We explain this in our article on filler rate. So the true share of "so" as a filler is lower than 49 percent. The direction of the finding holds: the speakers in this set used "um" rarely, and whatever they used instead was a common word.

Most speakers speed up at the end

For the 562 recordings longer than a minute, we compared the pace in the first third of the talk with the pace in the last third. The median rose from 130 to 135 words a minute, and 57 percent of speakers were faster at the end than at the start.

This is the opposite of what a good close usually needs. The last lines are the ones the audience is most likely to remember, and they are the ones most often rushed.

Pressure adds fillers, not speed

In live battles, where a speaker has a topic, a clock and an opponent, the median filler rate was 2.2 percent, against 1.7 percent in solo practice. That is about 30 percent more. Pace barely changed: 131 words a minute in battles, 128 alone.

In other words, under pressure these speakers did not rush. They reached for "so" and "like" more often while they thought.

What we cannot tell you yet

We wanted to show how much people improve with practice. We cannot, honestly, yet. Only a small number of speakers in this set have made ten or more recordings. Most of them had a higher overall score in their later recordings than in their first three, but only about half used fewer fillers, and a group that small is too few to say anything firm. We will publish that comparison when the group is large enough to mean something.

The other limits are worth stating plainly. These are people who chose to join a public speaking community, so they are not a sample of everyone. The recordings are short practice talks, not speeches to a real audience. And every number depends on how our instrument measures, which we publish in full in the How We Measure series, including the places where it gets things wrong.

If you want to use these numbers

You are welcome to quote them. Please link back to this page so readers can see the method and the limits alongside the figures.

The Exprea practice recorder and its measured analysis path: word-level timestamps from speech recognition, with filler words, pauses and speaking rate computed from those timestamps.

Measured

• Speaking rate in words per minute, per recording and across each recording in ten-second steps
• Every detected filler word, with its time
• Every silence longer than half a second
• Whether a recording was solo practice or a live battle turn

Inferred, not measured

• Whether a given "so" or "like" was a filler or a meaningful word: the instrument counts all of them
• Improvement over time: the group with ten or more recordings is too small to draw a conclusion