Tonality: measured, then judged
By Noam Shemla · Data & Method
TL;DR Tonality is vocal variety: how much your pitch moves, whatever its height. Your pitch is estimated in-browser by autocorrelation, four times a second, and summarised as a mean and a standard deviation in hertz. The 0–100 score itself is assigned by a model that has read those statistics, corroborated by your pace and by what other members voted.
Tonality is about variation, not about how high or low your voice is. Mean pitch is largely anatomy; movement around it is behaviour, and behaviour is the only thing worth scoring.
The axis used to be called Pitch on the AI side while members voting in Battles were already scoring something called tonality. Two names for one idea, drifting apart in the data. It is one name now, and the vote and the model land on the same vertex.
Where the numbers come from
In-browser Web Audio analysis, 2048-sample window, sampled 4× per second of recording; time-domain autocorrelation with parabolic interpolation, results outside 70–500 Hz discarded
Measured
- Mean pitch, Hz
- Standard deviation, Hz
- The monotone flag
- Strategic pause locations
Inferred, not measured
- The Tonality score, 0–100
The one part of this axis that is a hard rule is the monotone flag, and it is a ratio rather than a raw figure:
The strategic pause
A silence longer than 1.5 seconds that is not adjacent to a filler word is recorded as a strategic pause, and it nudges this axis upward: the power of silence, made countable.
A silence of the same length sitting next to an um is recorded as hesitation instead, and it costs you on Clarity. The same duration of nothing, read two ways, entirely on the basis of what was next to it.
This is a crude heuristic and it is only available on the measured path, since it needs word-level timestamps. It is an adjacency test and nothing more; it cannot judge whether a pause landed.
What pitch does to a listener
Pitch is the acoustic parameter most consistently tied to how a speaker is perceived, as distinct from how well they are understood.
Listeners judge confidence substantially from prosody rather than words. A speaker can say something certain and be heard as uncertain if the pitch contour rises where it ought to fall (Jiang and Pell, 2017).1 Work on student presentations found that measures of pitch variation corresponded to listener ratings of liveliness, giving a direct link between the statistic we compute and something an audience reports feeling (Hincks, 2005).2
There is also a well-known finding that listeners prefer lower-pitched voices when selecting leaders, for both men and women (Klofstad, Anderson and Peters, 2012).3
Read that last one carefully It concerns mean pitch, which is mostly anatomy, and it is a finding about listener bias, offered as information rather than advice. We do not score mean pitch and we would not suggest anyone try to change theirs. What this axis scores is variation, which is a habit and can be practised.
Blind spots in the pitch figures
It reports one number for the whole clip. A speaker who is expressive for twenty seconds and flat for forty can produce the same standard deviation as one who is moderately varied throughout.
It cannot tell useful variation from useless variation. Emphasis on a key word and an upward inflection at the end of every sentence both raise the standard deviation. Only one of them helps you.
It can be computed from very few samples. A pitch block is produced from as few as three valid frames, and the figure you are shown carries no indication of how many. Quiet speakers, whispered passages and unvoiced consonants contribute nothing, because they fail the loudness gate or fall outside the accepted range.
Often there is no measurement at all. The pitch statistics only survive on the measured path. If your recording took the estimated path (your microphone's noise floor can decide that for you), the model is still asked for a Tonality score, with nothing measured behind it. The score looks identical either way.
The captured timeline stops after roughly four minutes. Only the first thousand samples are kept, so on a longer recording everything past about 250 seconds contributes nothing to your pitch statistics.
And the score sits on top of the measurement. The gap between the two is widest on this axis, because the measurement is unusually good and the score sitting on top of it is a model's read.
Getting more movement into your voice
Reading aloud from material with dialogue in it, which forces pitch movement structurally rather than by intention. Speakers who try to "be more expressive" by will typically produce more volume rather than more pitch, which moves the wrong axis.
The other reliable route is landing the ends of sentences. A great deal of apparent flatness is every sentence ending on the same rising note, which reads as tentative and, statistically, as low variation.
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
- 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
- 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