What is vocal energy?
By Noam Shemla · Measurements
Vocal energy is a loudness figure derived from the raw waveform while you record. We report its average, minimum and maximum over a clip, on a scale from 0 to 1.
The short version This is the weakest measurement we take, and we would rather say so than let it be read as physics. The figure is a root-mean-square amplitude, multiplied by a fixed gain and clipped at 1, from an uncalibrated microphone, with no decibel scale behind it. It is useful for comparing moments within one recording and close to meaningless between recordings.
From waveform to a number
The browser reads the waveform as 8-bit samples, centres them, squares them, averages, and takes the square root. That gives a root-mean-square amplitude, the standard way to get a single loudness figure from a window of audio. That value is then multiplied by 3.2 and clipped to a maximum of 1.
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 energy across the clip
Inferred, not measured
- The Energy score (0–100), assigned by a language model from those three numbers
Each reading covers the last 2,048 samples of audio at the moment it is taken, the same window the pitch estimate uses. The 8-bit step is worth knowing about: it rounds every sample to one of 256 levels before anything else happens, which costs a little precision at the quiet end. The same readings are also averaged into 10-second slices for the energy timeline, and drawn as the waveform on your report's player, where each of its 80 bars shows the loudest reading in its stretch of the recording, scaled against the loudest moment of the take.
A worked example of the clipping. A passage whose raw RMS is 0.2 reads as 0.64. A passage at 0.35 would read as 1.12, and is clipped to 1. So would one at 0.5. Above a raw level of about 0.31, every louder passage looks the same.
Why the number does not travel
Every step between your voice and this figure is uncalibrated.
Microphone sensitivity varies by an order of magnitude across consumer hardware. Distance from the microphone matters a great deal: in open air the intensity of sound falls with the square of the distance, so a small lean back is a large change at the microphone. The room adds its own reflections. None of this is corrected for, because none of it is knowable from inside a web page.
Automatic gain control used to be on that list without qualification. It is the browser feature that raises quiet input and lowers loud input so a call sounds level, which is the opposite of what this measurement needs. Scored recordings now ask the browser to switch it off, while leaving noise suppression and echo cancellation on, because those help the transcription and do not flatten loudness in the same way. That request is a preference rather than a guarantee, and whether a given browser and device honour it is outside our control.
What this means in practice The same performance, recorded on a laptop built-in microphone and then on a headset, will produce different energy figures. If your Energy score moves between sessions, check whether anything about your setup moved first.
What it is good for
Within a single recording, the capture chain is constant, so relative energy is informative. A clip whose maximum is close to its average was delivered flat. One with a wide spread had dynamics in it: emphasis, contrast, a quiet line before a loud one.
That within-clip range is the sound use of this parameter, and it is what the coaching model is asked to read it for. The model receives the average, minimum and maximum together with the 10-second timeline, so it can see not only how wide the range was but when the louder and quieter stretches came.
The energy scale
- 0–1 scale. Dimensionless. Not decibels, not SPL, not LUFS.
- ×3.2 fixed gain. Applied before clipping, so loud passages saturate.
- 0.06 too-quiet warning. Peak below this after 2.5s triggers a recording warning.
The energy scale, in full.
The warning is the only place the raw figure is used as an absolute threshold. If, two and a half seconds into a recording, nothing louder than 0.06 has arrived, the screen tells you it can barely hear you and suggests moving closer or speaking up. It is a check that the microphone is working, not a judgement of your delivery.
Where it goes next
The Energy score for a recording comes from the coaching model, anchored to these readings. On a video take, a separate judgement of your on-camera presence can be blended in, at no more than 30% of the score, and less or not at all when the model reports low confidence in what it could see. That recording score then joins your pace, at a small weight, and peer votes from battles and rounds on the Energy axis. Energy: our weakest number covers the axis as a whole.
Where energy misleads
It does not measure how loud you were, in any unit a physicist would accept.
It does not survive a change of hardware, and nothing in the product flags when your hardware has changed.
It is nonetheless fed to the coaching model as an objective metric and contributes to an overall score that is compared against other members in the community benchmark. That comparison is between people on different microphones in different rooms. We are stating this plainly rather than quietly, because it is the kind of thing a measurement product should disclose about itself.
Loudness is not a fixed property of a voice. Acoustic work on vocal effort shows level, spectrum and pitch move together as a speaker pushes harder, and differently for men, women and children1; measurements of speakers and singers put the usable range far wider than conversation ever visits2. A single number for “energy” compresses all of that, which is why we report it as a relative curve and not in decibels.
Getting a fair reading
Record in the same place, on the same microphone, at the same distance, and compare takes only with each other. Then work on range rather than level: a deliberate drop before your key line and a clear lift on it will show in the spread, where speaking louder throughout will not. More on that in loudness is a choice.
References
- Traunmüller, H., & Eriksson, A. (2000). Acoustic effects of variation in vocal effort by men, women, and children. The Journal of the Acoustical Society of America, 107(6), 3438–3451. https://doi.org/10.1121/1.429414
- Titze, I. R., & Sundberg, J. (1992). Vocal intensity in speakers and singers. The Journal of the Acoustical Society of America, 91(5), 2936–2946. https://doi.org/10.1121/1.402929