Energy: our weakest number
By Noam Shemla · Data & Method
TL;DR Energy is a loudness figure from your microphone, scaled and clipped to a 0–1 range. It is not decibels and it does not survive a change of hardware. Within a single recording it tells you something real about your dynamics. Between recordings, or between people, it mostly tells you about equipment.
Every measurement product has a parameter it would quietly rather not discuss. This is ours. We are putting it in writing because the alternative, letting a number that cannot travel be read as if it can, is worse than the limitation itself.
Energy is meant to capture vocal commitment: whether you were audibly invested in what you were saying, and whether the delivery had dynamics in it or ran at one level throughout.
From microphone to score
In-browser Web Audio analysis, 2048-sample window, sampled 4× per second of recording; 8-bit time-domain RMS × 3.2, clipped to [0, 1]
Measured
- Average, minimum and maximum loudness over the clip
Inferred, not measured
- The Energy score, 0–100
The measurement itself is described step by step in what is vocal energy. In short: four times a second, the browser takes the last 2,048 samples of audio, computes their root-mean-square amplitude, multiplies by 3.2 and clips at 1. The coaching model receives the average, minimum and maximum of those readings and a 10-second timeline of them, and scores the recording from 0 to 100.
Two details the chart compresses. Members now rate a speaker's energy directly, on a scale of one to five that is converted to 0 to 100, and that rating carries the 0.6 weight; "body language" was the question older battles asked, and those votes still count on this axis. The overall-impression row applies only to those older votes.
Which sources count, and how much
Weights are only half the arithmetic. Each source is also multiplied by how far it can be trusted on this axis, and by a recency decay that halves every sixty days.
A recorded practice take with full measurement is trusted completely on Energy. A take that fell back to the estimated path, with no word timestamps, is trusted at 0.6. A Live Battle turn is trusted at 0.75, above an ordinary estimated take, because a timed turn against an opponent is a more telling sample of energy than a solo recording. Peer votes are trusted at 0.4. Live sessions contribute nothing here at all: they produce a transcript and no audio, and a transcript cannot speak to loudness.
So a peer's energy vote of four out of five, which is 80, enters with a weight of 0.6 times 0.4, or 0.24. A measured practice take scored at 70 by the model enters at 1.0. Recordings with fewer than thirty words are left out of the axis entirely.
Video takes
On a take recorded with the camera on, a separate judgement of your visible presence can be blended into that recording's Energy score. Its share is capped at 30%. If the model reports medium confidence in what it could see, the share drops to 18%; if it reports low confidence, because the frame was dark or you were barely in it, the blend is skipped and the vocal score stands.
A worked example: a vocal score of 60 and a presence score of 80, seen with high confidence, blend to 66. With medium confidence they blend to 64. With low confidence the score stays at 60. This is why an Energy score from a video take and one from an audio take do not mean quite the same thing, and the gap is bounded on purpose.
Why the number does not travel
Every stage between your voice and this figure is uncalibrated, and the browser is not neutral about it.
Automatic gain control deserves singling out. Beyond adding noise, it opposes what we are trying to measure, raising quiet input and lowering loud input. A speaker who deliberately drops to a near-whisper for effect may register almost no change at all. Scored recordings now ask the browser to turn it off, which is the best a web page can do. It is a request rather than a guarantee, so the summary above still holds for any browser or device that ignores it.
What it is good for
Within one recording the capture chain is constant, so relative energy is informative and the range is the useful statistic. A clip whose maximum sits close to its average was delivered flat. One with a wide spread had contrast in it: emphasis, a quiet line before a loud one, a change of gear.
That within-clip range is the honest use of this parameter, and it is what the coaching model is asked to read it for.
There is a second consolation, and a real one. A human ear is unaffected by your microphone, so peer votes are the one input to this axis that is immune to the problem the rest of it has. They are trusted less than a measured recording, but they are measuring something a microphone cannot distort.
How much the research backs it
Less than for the other axes, and it is better to say so than to pad.
Vocal intensity is one of several acoustic cues listeners draw on when judging how engaged or dominant a speaker is, and it moves with emotional arousal. But the findings that exist are about measured intensity under controlled recording conditions, and what we have is an uncalibrated proxy for intensity. Citing that literature as support for this score would be borrowing credibility the measurement has not earned.
If your Energy score moved Check whether your setup moved first. A different microphone, a headset instead of a laptop, or simply sitting further back will change this number more than any change in delivery will. Nothing in the product flags when your hardware has changed.
Where the figure misleads
Everywhere the number leaves the recording it came from. It is nonetheless fed to the coaching model as an objective metric, and it contributes to an overall score used in the community benchmark, which compares people on different microphones in different rooms.
We are stating this plainly rather than quietly because a measurement product's disclosures should be as easy to find as its numbers.
Improving your dynamics
Within a session, dynamics rather than volume. Speaking louder throughout raises the average and does nothing for the range, which is the part that carries. Dropping deliberately before a key line and then landing it does the opposite, and is also simply better speaking.
If you want the number itself to be comparable over time, the practical answer is unglamorous: use the same microphone at the same distance every time, and treat any jump on this axis after a hardware change as noise. And if the votes from a battle say something different from your recordings, take the votes seriously; on this axis they are the reading least affected by equipment.