Methodology

1. Signal preparation

Your recording is decoded in the browser, mixed to mono and resampled to 22.05 kHz 16-bit WAV. The server re-validates format, length (1.5 s – 5 min), level and clipping.

Browser voice processing (noise suppression, auto-gain, echo cancellation) is disabled while recording because it alters timbre measurements.

2. Pitch tracking

Fundamental frequency is tracked every 10 ms with Praat's autocorrelation algorithm (Boersma, 1993) between 60 and 1200 Hz, covering bass to soprano. Voiced runs shorter than 50 ms are discarded.

Notes are segmented where the smoothed pitch moves more than 80 cents and stays there for 80 ms, which keeps vibrato from splitting notes. Range uses the median of each note, not single frames, so glitches and slides do not inflate it.

3. Stability and vibrato

Within each sustained note (≥ 0.3 s, edges trimmed) pitch is split into a slow trend (250 ms moving average) and a fast residual. Drift is the spread of the slow trend; the residual spectrum is searched for a 3.5–8.5 Hz vibrato peak, and the remaining non-vibrato energy is reported as irregular fluctuation.

Vibrato is only called when its semi-extent reaches 45 cents. This threshold was chosen on 17 professional singers (VocalSet): it detects 84% of takes sung with vibrato and wrongly flags 8% of straight-tone takes (a 12-cent threshold flagged 76%). The stability score maps the combined deviation from 5 cents (100) to 50 cents (0); straight long tones of professionals score a median of 88.

4. Voice-quality measurements

HNR (harmonics-to-noise ratio, cross-correlation method), CPPS (smoothed cepstral peak prominence; Hillenbrand & Houde, 1996), local jitter and shimmer are computed with Praat on voiced segments only. H1–H2 is measured from the spectrum at f0 and 2·f0 and flagged unreliable above ~350 Hz.

Spectral balance uses the long-term average spectrum of voiced frames: alpha ratio (1–5 kHz vs 0.05–1 kHz), Hammarberg index and singing power ratio (2–4 kHz vs 0–2 kHz peak; Omori et al., 1996).

5. Tone-quality estimates

Breathiness, clarity, brightness and resonance are ESTIMATES calibrated on 17 professional singers (VocalSet technique labels). Each acoustic feature passes through a logistic curve centred between the class medians, so about 40 = a professional's straight tone and 60+ = a professional's deliberately breathy tone, without saturating for amateur voices.

Breathiness uses HNR, CPPS and H1–H2 (jitter and shimmer were excluded because they marked healthy vibrato as breathy). Roundness measures acoustic rounding - a lowered second formant (F2) and a darker spectral centre - calibrated on rounded (o, u) versus spread (i, e) vowels of the same singers: AUC 0.93, where the earlier heuristic scored 0.52 (chance). F2 is skipped for very high voices. Thinness combines the opposite of roundness (60%) and of resonance (40%). Microphone, room, vowel and pitch still shift all of these, so compare recordings made under similar conditions.

6. Pitch accuracy

Pitch accuracy is only computed when target notes exist. Sung notes are aligned to the targets in order (dynamic programming that tolerates missed or extra notes), or by time window when timing is given.

Each note scores 100 within ±10 cents, falling linearly to 0 at ±100 cents. 'In tune' means within ±25 cents.

7. Overall score

The overall score blends pitch stability (30%), pitch accuracy against target notes (35%) or, without targets, reference-free intonation (25%), clarity (20%) and resonance (15%). Pitch control is weighted most because on 575 real singers rated by trained listeners (SingMOS-Pro) stability (ρ +0.27) and intonation (ρ −0.21) tracked human ratings best, while clarity did not.

The scale is anchored to professional singers: 85 = a typical professional on VocalSet, 70–95 = the professional range (10th–90th percentile). Grade A means professional level. Agreement with human quality ratings is still modest (ρ ≈ 0.26), so the overall score remains an estimate and not a substitute for a vocal teacher.

Validation on real singers

VocalSet (Wilkins et al., 2018; CC BY 4.0): 2,313 recordings of 17 professional singers with technique labels. Separation, measured as AUC (0.5 = chance, 1 = perfect): breathiness 0.87 (was 0.69 before calibration), clarity 0.86 (was 0.75), brightness 0.81 (was 0.79). Vibrato: 84% detected, 8% false alarms (was 76%).

vocadito (Bittner et al., 2021; CC BY 4.0): solo singing by singers of varying training, with pitch and note annotations by musicians. On the 13 excerpts available to us: raw pitch accuracy 97.1% (frames within ±50 cents of the annotation), voicing recall 98.5%, 91–93% of annotated notes found, median note pitch error 2–3 cents.

SingMOS-Pro (Tang et al., 2025; CC BY 4.0): 575 recordings of real singers, each rated by at least five trained annotators. Spearman correlation with the human quality score: overall +0.26 on datasets not used for tuning, stability +0.27, intonation deviation −0.21.

Roundness: AUC 0.93 separating rounded from spread vowels (209 takes, 17 singers). Not yet validated: the overall score as a judgement of artistry, and the scale for beginners (all calibration singers were professional or semi-professional). Benchmark scripts: vocaltone-ai/benchmarks/.

Limitations

Consumer microphones, room echo and background noise affect HNR, CPPS and spectral measures. Jitter and shimmer are inflated by vibrato. Very high voices (above ~500 Hz) make spectral-envelope measures less reliable, which lowers the reported confidence.

VocalTone is a practice and self-monitoring tool. It does not diagnose voice disorders.