What your voice says before you do

The engine is a from-scratch port of the core algorithms in Praat, the standard tool in phonetics research: Boersma’s autocorrelation pitch tracker, cross-correlation harmonicity, glottal-pulse marking for jitter and shimmer, and Burg linear-prediction formant tracking. It runs in your browser first, so the score is on screen at once. Then our server measures your recording again with the same engine and gets the same number to the last digit. A saved reading, a rating and a challenge link are all held to the server’s measurement. That is what “verified” means on your result.

We ran it against real Praat on 14 recordings of real voices, same file through both. Median pitch agreed within 0.4%, pitch variability within 0.08 semitones, jitter within 0.12 percentage points, shimmer within 0.3, and harmonics-to-noise ratio within 0.4 dB. Of 42,552 formant frames, all but 166 came out identical to Praat’s, and the formant medians we score from agreed within 0.1 Hz.

Men’s scores split the voice into the instrument you were given and the delivery you train:

Pitch. The most replicated result in the field: women rate lower-pitched men’s voices as more attractive. It stops paying off at the bottom: below about 96 Hz, women pick the higher voice. So nothing extra under 96 Hz, and don’t force yours down. A pushed-down voice sounds strained and more dominant, not more attractive.

Resonance. Formants are the resonances of your throat and mouth. Lower ones make a voice sound as if it comes from a bigger man. We measure the first four and report them as an apparent vocal-tract length: the tube that would produce them, not a measurement of your anatomy. The average man in our norm group comes out at 17.9 cm. It counts, for far less than pitch.

Projection. Men who speak louder are rated more attractive. A phone sets its own gain, so it can’t read loudness directly. But a louder voice shifts its energy above 1 kHz, and that balance is something gain can’t fake. So that’s what we read, in the pitched parts of your recording.

Modulation. When a man records a message for a woman he wants, his pitch moves more and his phrase endings drop lower. Women rate that voice higher, even when they don’t understand the language. What counts is whether you switch it on, so we don’t compare your melody with other men’s. You read the line twice, plain and then like you mean it, and we measure how much you changed.

Pace. Slow a man’s speech by 15% and women rate him less attractive; speed it up by 15% and it helps a little. We take points off for slow or halting delivery and give none for rushing. Everyone reads the same line, so pace is exact: known syllables over the time you spent speaking.

Smoothness. When 800 listeners compared the same speakers with and without vocal fry, the creaky version lost, for men and women alike. Creak costs points. Finer roughness (jitter, shimmer) gets the smallest weight.

Women’s scores use a simpler model:

Men rate higher-pitched women’s voices as more attractive, and the preference holds across the whole 160–300 Hz range.

What we don’t score. Timbre and airiness show as descriptions only. We don’t reward a “cleaner” voice: the one experiment that varied breathiness found the breathier version rated more attractive, so calling clarity better would be inventing a rule.

We ran this exact engine over 40 male and 40 female speakers from LibriSpeech, two takes each. A 5.0 is the median voice in that group; every 1.6 points is one standard deviation. The highest score anyone in it reached was 9.6. High scores are rare.

Two recordings of the same man never give identical scores. So everyone reads the same line, we show your margin (about ±0.4 points for a man’s reading), and extra readings are averaged rather than letting you keep your best one.

Smoothness is only scored in a quiet enough room: clear gaps between phrases, at least 20 dB between your voice and the room. Otherwise we leave it out and tell you, rather than hold your room against you.

Your recordings. Each take is uploaded after it is scored and kept, with the server’s measurement of it. We keep them for three things: to verify your score against the audio rather than take a phone’s word for it; so you can listen back to your own takes as you train; and to improve the model — the norms, the retake error and, once enough people have read the line, the weights themselves, re-fitted from real recordings and real listeners.

A recording is never used to identify you, and never sold or shared. With an account, your recordings are yours to play back and are deleted with the account. Without one, a recording carries no name — only the country it came from and a salted hash of the connection — and is destroyed after three years. In places where a voice recording counts as biometric data, we ask for your explicit consent before recording. The women’s panel plays a take to real people only with a separate consent, and that copy is deleted when the ratings are in.

The six-week target. It is this same model run on your pitch and resonance as measured, with every part of your delivery trained to one standard deviation above the average man. That is what six weeks of reps is aimed at.

A low score. The biggest single part of a man’s number is pitch, and a low score usually means yours sits higher than women prefer. That part is given. Delivery — projection, how you open up when you mean it, pace, smoothness — is 45% of the score, and it’s the part training moves.

Hearing yourself. Your recorded voice sounds different to you from the one in your head, because you hear yourself partly through the bones of your skull, which carry the low frequencies. Everyone else only ever hears the recorded version. That is the voice we measure, and the one you train.

• Anderson, Klofstad, Mayew & Venkatachalam (2014). Vocal fry may undermine the success of young women in the labor market. PLoS ONE 9(5).
• Babel, McGuire & King (2014). Towards a more nuanced view of vocal attractiveness. PLoS ONE 9(2).
• Boersma (1993). Accurate short-term analysis of the fundamental frequency and the harmonics-to-noise ratio of a sampled sound. IFA Proceedings 17.
• Collins (2000). Men’s voices and women’s choices. Animal Behaviour 60.
• Collins & Missing (2003). Vocal and visual attractiveness are related in women. Animal Behaviour 65.
• Feinberg, Jones, Little, Burt & Perrett (2005). Manipulations of fundamental and formant frequencies influence the attractiveness of human male voices. Animal Behaviour 69.
• Feinberg, DeBruine, Jones & Perrett (2008). The role of femininity and averageness of voice pitch in aesthetic judgments of women’s voices. Perception 37.
• Hodges-Simeon, Gaulin & Puts (2010). Different vocal parameters predict perceptions of dominance and attractiveness. Human Nature 21.
• Leongómez et al. (2014). Vocal modulation during courtship increases proceptivity even in naive listeners. Evolution and Human Behavior 35.
• Panayotov, Chen, Povey & Khudanpur (2015). LibriSpeech: an ASR corpus based on public domain audio books. ICASSP.
• Pisanski et al. (2016). Volitional exaggeration of body size through fundamental and formant frequency modulation in humans. Scientific Reports 6.
• Puts, Apicella & Cárdenas (2012). Masculine voices signal men’s threat potential in forager and industrial societies. Proceedings of the Royal Society B 279.
• Quené, Boomsma & van Erning (2021). Attractiveness of male speakers: effects of pitch and tempo. In Weiss et al. (eds), Voice Attractiveness. Springer.
• Re, O’Connor, Bennett & Feinberg (2012). Preferences for very low and very high voice pitch in humans. PLoS ONE 7(3).
• Reby & McComb (2003). Anatomical constraints generate honesty: acoustic cues to age and weight in the roars of red deer stags. Animal Behaviour 65.
• Sundberg & Nordenberg (2006). Effects of vocal loudness variation on spectrum balance as reflected by the alpha measure of long-term-average spectra of speech. Journal of the Acoustical Society of America 120(1).
• Xu, Lee, Wu, Liu & Birkholz (2013). Human vocal attractiveness as signaled by body size projection. PLoS ONE 8(4).