In short: Our voice tools listen through your microphone with the browser’s Web Audio API, find the pitch of each moment with the YIN method, turn it into a note with A4 = 440 Hz, and keep only notes you hold steadily. Your lowest and highest held notes give your range, octaves and closest voice type. Ten tools were tested before release: 285 checks passed, 0 failed, 50 could not be run.
The Measurement Chain
Step 1: Microphone Access
Nothing listens until you press Start. The browser then asks for permission, and the tool reads the signal through the Web Audio API. We ask the browser to switch off echo cancellation, noise suppression and automatic gain control, because they bend the signal, although some phones apply them anyway. Only one tool on a page holds the microphone at a time, and it is released when you stop, reset, finish or leave the page. If access is blocked, the tool switches to choosing notes by hand.
Step 2: Finding the Pitch
Each short block of sound (2,048 samples at your device’s real sample rate) goes through YIN, a published pitch detector (de Cheveigné and Kawahara, 2002), with a threshold of 0.15 and interpolation between samples. In the range tools, a reading is used only when it lies between 60 and 1,400 Hz, is clear enough (a periodicity score of at least 0.8) and loud enough, and the tool smooths the last five readings with a median.
We switched to YIN because the old detector often locked onto a multiple of the true period and reported notes an octave too low. On the same 216 synthetic test frames, from 82 to 1,047 Hz, the old range test got 108 wrong and the new engine none.
Step 3: From Hz to a Note
Notes follow the equal-tempered scale with A4 = 440 Hz: note number = 69 + 12 × log2(frequency ÷ 440). So 131 Hz is C3, 220 Hz is A3 and 523 Hz is C5. The tools use sharp names (C#), and the calculator also shows the flat name (Db) and each note’s frequency.
Step 4: Only Steady Notes Count
A breath, a slide or a moment of noise can produce a stray reading, so the range tools only offer a note after you have held it within one semitone for 0.8 seconds; vibrato still counts. A held note expires two seconds after you stop, and you choose when to save it, so the tool never keeps a note just because you started on it. Silence and white noise produce nothing to save; both are part of the tests.
Step 5: Range, Octaves and Voice Type
Range in semitones is highest minus lowest; octaves are semitones ÷ 12. C3 to C5 is 24 semitones, exactly 2 octaves; F3 to C6 is 31 semitones, 2.58 octaves (2 octaves plus a perfect fifth). The voice-type tools compare your range with the ranges on our own voice type pages:
| Voice type | Range used | Voice type | Range used |
|---|---|---|---|
| Soprano | C4–C6 | Tenor | C3–C5 |
| Mezzo-soprano | A3–A5 | Baritone | A2–A4 |
| Contralto | F3–F5 | Bass | E2–E4 |
The match is the overlap between your range and a type’s range divided by the two combined, in semitones. A small penalty for distance between the centers decides close cases, the result says “between X and Y” when two types fit almost equally, and a best match under 25% is reported as no close match. Range is only part of voice type, so we present it as a starting point.
Speaking Pitch and the Singing Score
- Speaking pitch: the Voice Frequency Test uses your median pitch over a 15-second reading and compares it with typical adult bands (about 90–155 Hz and 165–255 Hz, from Baken’s Clinical Measurement of Speech and Voice). It never guesses gender. The voicecel test places the same median on a five-step scale, from very deep (under 100 Hz) to very high (220 Hz and above).
- Singing score: AI Voice Analysis and Rate My Voice share one scoring engine. It splits your take into held notes and weighs pitch accuracy (40%), steadiness (25%), share of held notes (15%), range (10%) and clarity (10%). Accuracy is measured against your own tuning, so a singer who is evenly flat isn’t penalised; about 25 cents off is where listeners start to hear a note as out of tune. The rules are fixed, not a trained AI model, and so far they have been checked on synthetic voices only.
How Each Tool Was Tested
Every rebuilt tool was installed on a clean WordPress test site and driven in a Chromium browser with a simulated microphone playing generated audio: held notes, glides with vibrato, synthetic speech with a known pitch, silence and noise. The same tests covered manual input, error messages, phone-width layouts, and whether the microphone is released.
| Tool (live version) | Key check | Passed | Failed | Not tested |
|---|---|---|---|---|
| Vocal Range Test (2.0.0) | New pitch detector: 0 of 216 synthetic test frames wrong; the old 1.0.0 detector got 108 of 216 wrong (50%), including octave errors | 30 | 0 | 6 |
| Voice Frequency Test (1.0.0) | Synthetic speech with known medians of 119.4 Hz and 208.9 Hz measured with 0.3% and 0.5% error | 24 | 0 | 5 |
| Voice Type Test (2.0.0) | Each of the six voice-type ranges returns its own type at a 100% match; A2–E4 returns “between baritone and bass” | 34 | 0 | 5 |
| Voice Type Checker (on the voice-type pages) (1.0.0) | E2 (82 Hz) with a strong second harmonic read as E2, not E3; C6 (1,047 Hz) detected; keyboard tab order checked | 35 | 0 | 5 |
| Vocal Range Calculator (2.0.0) | F3–C6 = 31 semitones, 2 octaves + perfect 5th; C3–C5 = 24 semitones, frequency ratio 4.00 : 1 | 31 | 0 | 4 |
| Octave Range Test (2.0.0) | F3–C6 = 2.58 octaves with octave-by-octave coverage; microphone glide A2→E4 = 1.58 octaves | 24 | 0 | 5 |
| Singer Comparison Tool (3.0) | Top-8 singer matches identical to an independent Python calculation for three test ranges (181 singers) | 35 | 0 | 5 |
| AI Voice Analysis (2.0.0) | In-tune synthetic singing with vibrato scored 9.8 out of 10; notes 22–42 cents off scored 6.5 | 26 | 0 | 5 |
| Rate My Voice (2.0.0) | Scoring code identical to AI Voice Analysis (source comparison), so both give the same score | 23 | 0 | 5 |
| Voicecel Test (2.0.0) | Synthetic speech at about 118 Hz → “Deep” (119 Hz median); about 210 Hz → “High” | 23 | 0 | 5 |
| Total | 10 tools | 285 | 0 | 50 |
What We Could Not Test
- Real human voices: the tests used synthetic singing and speech. Your own test on the live site is the real check.
- Safari on Mac and iPhone, Firefox, and Chrome on a physical Android phone (a Safari fallback is written into the main tools but was not run).
- How each tool sits in the live theme with ads around it.
Some older tools have not been rebuilt yet: the deep voice test, note detector, pitch test, singing test, song key finder, vocal range finder, vocal tier test, voice pitch test and voice tone test. They work, but they don’t have test reports like the table above. We checked their scripts too: none of them sends audio anywhere.
Bugs the Tests Caught
- The old range test auto-saved the first note you held as your lowest, even if it was just your starting note. You now save each note yourself.
- AI Voice Analysis showed a “range used” that included a note never sung; it now uses your lowest and highest held notes.
- In the Singer Comparison Tool, choosing notes in reverse order corrupted the next choice.
- On a 360-pixel phone, the range test’s Start button sat below the first screen; the layout was changed.
Your Audio Stays on Your Device
All of the analysis above runs in your browser. The tools don’t record, store or upload audio, and they keep nothing in cookies or browser storage; a result disappears when you reload. Copying a result or saving the Rate My Voice score image happens on your device. See data security.
What the Tools Don’t Measure
They can’t judge vocal health, tone quality, stamina or whether a note is produced safely, and a range-based voice type is not a teacher’s classification. Never push for a note that hurts. How to get a steadier result is covered in how to find your vocal range; the limits are in our disclaimer.
Retesting and Corrections
Each new version of a tool gets a complete new test report, not only a check of what changed. A reported bug is first reproduced with a known signal, then fixed, and that check joins the test list. This page changes whenever the results table changes. Report a problem through the contact page; our wider standards are in the editorial guidelines.
Last updated: 27 September 2026