What is a pause?
By Noam Shemla · Measurements
A pause in our analysis is any gap between the end of one word and the start of the next that lasts longer than 0.5 seconds. There is no upper bound and no minimum count.
The short version Half a second is the threshold. It is a defensible choice rather than a discovered constant, and everything about the measurement follows from it. The instrument records where pauses happened and how long they were. It makes no attempt to say whether they were intentional.
Finding the gaps
Word-level timestamps give the end time of every word and the start time of the next. The difference is the gap. Every gap over the threshold is recorded with its position and its length, both rounded to hundredths of a second. The position is the moment the previous word ended, so a pause is pinned to the word it follows.
Deepgram nova-3, word-level timestamps with per-word confidence
Measured
- Pause position, to two decimal places
- Pause duration
Inferred, not measured
- Whether a pause was rhetorical or a stumble. Nothing in the pipeline attempts this
This measurement needs word timestamps, so it exists only when your recording went down the measured path, where the transcription service returned words with enough confidence. When a recording falls back to the estimated path, there are no pauses on the report at all, rather than pauses we guessed.
Two jobs a pause does
Pausing does two jobs at once, and they pull in the same direction.
The first is for the listener. Speech is processed in chunks, and a pause marks a boundary between them. Silence at a syntactic boundary helps comprehension; silence in the middle of a phrase hurts it. Where you pause matters more than how often.
The second is for the speaker's credibility. A pause before a key phrase is one of the few devices that reliably reads as control rather than hesitation. It signals that the next thing is worth waiting for. It is also the direct substitute for a filler: the same moment of thinking, made silent. There is more on that trade in what um actually tells your audience.
The pause specification
- 0.5s threshold. Gaps above this are recorded as pauses.
- none upper bound. A ten-second silence is one pause, same as a one-second silence.
- 2 dp timing precision. Positions and durations are rounded to hundredths of a second.
The complete specification of the pause measurement.
Long pauses get a second look
The list of recorded pauses is handed to the coaching model as context. On top of that, pauses of 1.5 seconds or longer are sorted into two kinds by a simple rule. If a filler word falls within 0.3 seconds of either edge of the silence, the pause is counted as a hesitation. If not, it is counted as a strategic pause.
That rule is an adjacency test, not an understanding of the talk. A silence followed by um is treated as a speaker searching; a clean silence is given the benefit of the doubt. A long silence in which you lost your place entirely, then resumed without a filler, will be counted as strategic. Pauses between half a second and 1.5 seconds are recorded but not classified at all.
Where pauses enter the score
When a measured recording contains at least one strategic pause, it adds a contribution worth 75 out of 100 to Tonality and to Structure. When it contains at least one hesitation, it adds a contribution worth 35 to Clarity. Each carries a weight of 0.15, small beside the coaching model's own score on that axis, which carries 1.0. It is the presence of each kind that counts, not the number: three strategic pauses add the same as one. There is more on how these weights combine in how Exprea scores a speaker.
What silence data cannot show
It cannot distinguish intention. This is the measurement's central limitation and no amount of tuning fixes it: a deliberate beat before a punchline and a speaker searching for a word produce identical data. The 1.5-second classification narrows the problem without solving it.
It does not see pauses at the edges. The gap before your first word and after your last are not inter-word gaps, so a long, uncomfortable silence at the start of a talk is not recorded as a pause at all. It does still count against your speaking rate, because that divides by the full clip length.
It does not know about breath. Many of the gaps a listener hears as pauses are shorter than the threshold, and are invisible here.
Why half a second
Half a second is our own line, and the research gives good reason to say so rather than dress it up. Frieda Goldman-Eisler's work on spontaneous speech found that its pauses follow its syntax, with the larger structures marked differently from the smaller ones, and that reading the same words aloud changes where the pauses fall.1 Pauses are part of how speech is planned, not just gaps in it.
Estelle Campione and Jean Véronis then measured about 6,000 silent pauses in five and a half hours of read and spontaneous speech in English, French, German, Italian and Spanish.2 On a logarithmic scale the durations fell into three groups: brief pauses under 200 milliseconds, medium ones from 200 to 1,000, and long ones over a second, which appeared only in spontaneous speech. They also warned that fixed thresholds, used widely for convenience, can produce false conclusions when comparing speakers, styles or languages.
Our 0.5 seconds sits in the middle of their medium group. It ignores every brief pause and part of the medium range. That is a reasonable cut for coaching, because the pauses a listener notices are mostly longer than that, but it is a convention, so it is stated rather than justified.
Using the pause data
Look at where your pauses fall, not how many. A pause after a sentence is doing its job; a run of them inside one sentence is the place to rehearse.
Try putting one deliberate silence of a second or more before your main point, with nothing on either side of it. If your report lists a hesitation, find the filler beside it: that is the moment to replace the sound with the silence. More on the craft side is in the pause is punctuation.
References
- Goldman-Eisler, F. (1972). Pauses, clauses, sentences. Language and Speech, 15(2), 103–113. https://doi.org/10.1177/002383097201500201
- Campione, E., & Véronis, J. (2002). A large-scale multilingual study of silent pause duration. Speech Prosody 2002. https://doi.org/10.21437/speechprosody.2002-35