What is pitch variation?
By Noam Shemla · Measurements
Pitch variation is how much your fundamental frequency moves while you speak. We report it as a standard deviation in hertz, alongside your mean pitch, and derive a monotone flag from the ratio between them.
The short version Pitch is estimated in your browser, four times per second, by autocorrelation over a 2048-sample window. Any result outside 70–500 Hz is thrown away. If the standard deviation divided by the mean falls below 0.10, we call the delivery monotone. The measurement is real; the Tonality score derived from it is a model judgement.
Measuring pitch in the browser
Nothing leaves your machine for this one. While you record, the browser samples the raw waveform and estimates the fundamental frequency by autocorrelation: sliding the signal against itself and finding the lag at which it best matches, which is the period of the voice. A parabolic interpolation refines that lag to sub-sample accuracy.
In-browser Web Audio analysis, 2048-sample window, sampled 4× per second of recording; time-domain autocorrelation with parabolic interpolation
Measured
- Mean pitch in hertz
- Population standard deviation in hertz
- The monotone flag
Inferred, not measured
- The Tonality score (0–100), which a language model assigns after reading these statistics
Each reading is a snapshot: the most recent 2,048 audio samples at the moment of the tick, not an average of the quarter second since the last one. Averaging would smooth away exactly the movement this parameter exists to catch.
The 70 to 500 Hz band works as a filter on the answer rather than a limit on the search. Every lag is tried, the best one wins, and if the frequency it implies falls outside the band, that reading is discarded. Below 70 Hz and above 500 Hz the estimator is usually locking onto harmonics or room noise rather than a voice. Very quiet frames are dropped before any of this, so whispering and the tail of a fading word contribute nothing at all.
The monotone rule
Monotone is not a judgement here; it is a threshold. We divide the standard deviation by the mean (a coefficient of variation) and call anything below 0.10 monotone.
A worked example. A speaker averaging 120 Hz with a standard deviation of 11 Hz has a ratio of about 0.09 and is flagged monotone. A speaker averaging 220 Hz with a standard deviation of 20 Hz has a ratio of about 0.09 too, and is flagged the same way, even though their voice moves nearly twice as many hertz. Raise the first speaker's spread to 15 Hz and the ratio is 0.125: not monotone.
Separately, the report describes anything at or above 30 Hz of standard deviation as Expressive, a flagged delivery as Monotone, and everything in between as Some variation. That is a display band only and plays no part in the score. It is also the one place where raw hertz is used, so it is easier for a higher voice to be called expressive than a lower one moving by the same proportion. The monotone rule avoids that problem; the label does not.
How listeners hear pitch
Pitch is the acoustic parameter most consistently linked to how a speaker is perceived.
Two findings hold up reasonably well. Listeners judge confidence substantially from prosody rather than words: a speaker can say something certain and be heard as unsure if the pitch contour rises where it should fall (Jiang and Pell, 2017).1 And in a voting context, listeners preferred candidates with lower-pitched voices for both men and women, a preference visible in experimental work on leadership selection (Klofstad, Anderson and Peters, 2012).2
Read that second finding carefully It is about mean pitch, which is largely anatomy, and it is a finding about listener bias, offered as description rather than advice. We do not score mean pitch and we do not suggest anyone try to change theirs. What we score is variation, which is behaviour and can be practised.
The link between variation and the audience is the one that matters here. Work on student presentations found that measures of pitch variation corresponded to listener ratings of liveliness (Hincks, 2005).3 That is the closest thing in the literature to a direct check on the statistic we compute, and it is why we use a ratio to the speaker's own mean rather than a fixed number of hertz.
Pitch settings
- 70–500 Hz search band. Estimates outside this are discarded as noise.
- 4 / sec sampling rate. Of recording time; paused time is excluded.
- 0.10 monotone threshold. Standard deviation ÷ mean.
The complete pitch measurement specification.
Where it goes
The coaching model reads your mean, your standard deviation, the monotone flag and a pitch timeline averaged over 10-second slices, and gives a Tonality score for the recording. That score enters the Tonality axis at a weight of 1.0. Your pace adds a smaller contribution at 0.25, a strategic pause at 0.15, and peer votes from battles and rounds add theirs at 0.6, scaled down by a lower trust in peer judgement.
Limits of a single standard deviation
It does not tell you where the variation was. A speaker expressive for twenty seconds and flat for forty can produce the same standard deviation as one moderately varied throughout. The timeline helps; the headline number does not.
It does not distinguish variation that helps from variation that does not. Emphasis on a key word and an upward inflection at the end of every sentence both raise the standard deviation. Only one of them is doing you a favour.
It cannot tell a voiced sound from a noisy one. The estimator has no voicing decision, so breath, an s or room noise can produce readings that widen the spread. We record how periodic each frame was, so a filter can later be chosen from real recordings; none is applied yet.
It uses a population standard deviation, which slightly understates the spread on very short clips.
And the score sits one step away from the measurement. The 0–100 Tonality number is assigned by a language model that has been told your standard deviation and your monotone flag, so it is anchored to those figures without being computed from them.
Practising variation
Pick the three words that carry your point and give each a clear change: higher, lower or slower than the words around it. Then look for the flat stretch on the timeline. More on what flatness costs is in vocal variety.
References
- Jiang, X., & Pell, M. D. (2017). The sound of confidence and doubt. Speech Communication, 88, 106–126. https://doi.org/10.1016/j.specom.2017.01.011
- Klofstad, C. A., Anderson, R. C., & Peters, S. (2012). Sounds like a winner: voice pitch influences perception of leadership capacity in both men and women. Proceedings of the Royal Society B, 279(1738), 2698–2704. https://doi.org/10.1098/rspb.2012.0311
- Hincks, R. (2005). Measures and perceptions of liveliness in student oral presentation speech. System, 33(4), 575–591. https://doi.org/10.1016/j.system.2005.04.002
Go deeper: Tonality: measured, then judged