About Voice Rater

A free, independent voice analysis tool built on one principle: measure the voice honestly, and never take the audio off your device.

Why this site exists

There are plenty of places on the web that will rate your voice. Looking closely at how they work, most fall into one of two groups. Some upload your recording to a server for processing, which means your audio leaves your control entirely. Others do not appear to measure anything at all, and hand back a number that changes on every attempt regardless of what you said.

Voice Rater exists because there is a third option that nobody was offering: run genuine acoustic analysis locally, in the browser, and show the user every number behind the result.

What it actually measures

After you stop recording, your audio is split into 40-millisecond frames on your own device. Silent frames are discarded. Every remaining frame is analysed with the McLeod pitch method, a normalised autocorrelation, to extract your fundamental frequency. The tool then derives:

• Smoothed cepstral peak prominence, or CPPS, the primary voice quality measure. Each frame's log power spectrum is transformed again to give a cepstrum, and the height of the peak at your pitch period above the baseline says how closely your spectrum resembles a clean harmonic comb. It is the standard measure for connected speech.
• Jitter, cycle-to-cycle variation in pitch period, measured in the time domain by locating individual vibration cycles.
• Shimmer, the same variation applied to amplitude.
• Signal-to-noise ratio, your speaking level against the noise floor in the silent frames, which is how the tool judges whether the recording was clean enough to trust.
• Harmonic-to-noise ratio, derived from the normalised autocorrelation peak, now mainly a fallback for takes too short or broken for a cepstral estimate.
• Spectral centroid, computed with an FFT, describing tone brightness.
• Formants F1 and F2, extracted through LPC spectral envelope analysis on a downsampled copy of the signal.

These are the same parameters used in speech science and voice clinics. They are combined into the 1-10 score with fixed, published weights: clarity 32%, pitch stability 16%, expressiveness 14%, volume control 12%, tone brightness 12%, delivery flow 8%, range 6%. The full weight table is published on the how it works page, alongside the threshold each sub-score is measured against, so anyone can check what drove their result.

How the score is calibrated

A scoring curve is a design decision, and it is easy to get wrong in both directions. Score everyone at nine and the number carries no information. Score everyone at four and the tool is simply rude.

The curve here is calibrated on measured recordings of people reading aloud, not on clinical reference values, because those are defined on a sustained vowel and read far higher in normal speech. It is set so that a healthy adult voice recorded on ordinary consumer hardware lands in the sevens, a trained voice in a quiet room reaches the nines, a voice with a heavy cold falls to around five and a half, and genuinely disordered voice measurements drop near three. The result is deterministic: record the same passage twice in the same room and the score repeats to within a few tenths.

Your audio never leaves your device

Every calculation happens in your browser through the Web Audio API. There is no upload, no server-side processing, no account and no storage. Your recording exists only in the memory of your browser tab and is discarded the moment you record again or close it.

This follows from how the site is built. It is a set of static files with no backend that could receive audio, so the guarantee does not depend on a policy.

Who runs it

Voice Rater is built and maintained by Daniel Hart, an independent developer working on small, focused web tools. There is no company behind it, no investors, and no data business. The site is free to use, with no premium tier and nothing gated behind a sign-up.

The tests behind the scoring are written up in the guides, each one measured on real public-domain recordings with the method shown, so the claims on this site can be checked rather than taken on trust.

Changelog

The tool changes when a measurement turns out to be wrong. The main changes so far:

• August 9, 2026. Launched with the speaking, singing and gender tests, all analysis running in the browser.
• August 2026. Jitter and shimmer moved to true cycle-by-cycle measurement after the first frame-based version understated them by about ten times.
• September 6, 2026. Scoring recalibrated on real human readings instead of synthetic test tones. Cepstral peak prominence became the main clarity measure, and recording-quality compensation was added so a noisy room no longer makes a voice read older.
• September 18, 2026. Example results from a public-domain reading added to every test page, and the 255 Hz boundary between pitch bands made consistent across tools.
• September 21 to 23, 2026. Result cards and sub-score bars can be opened for a plain-language explanation of each number.
• October 1, 2026. New design, and the guides section with three measured tests: take-to-take variation, real voices versus synthetic tones, and the effect of background noise.
• October 2, 2026. An optional deep resonance check on the gender detector, a way to save readings and track progress in your own browser, and three more guides measured on 40 LibriSpeech readers: pitch versus resonance, male and female voice frequency, and tracking voice training.
• October 3, 2026. Japanese versions of the home page, voice age test and gender detector; a vocal range test that finds your lowest and highest steady notes and your closest voice type; a playable public-domain sample on the home page; and clearer help when the microphone is blocked or silent.

What this site is not

It is not a medical device and not a diagnostic service. Cepstral peak prominence, jitter and shimmer are used in voice clinics, where they are measured on calibrated equipment under controlled conditions and interpreted by a clinician alongside a physical examination. A five-second recording on a laptop is not equivalent, and no result here should be treated as a health assessment. If your voice is persistently hoarse for more than two or three weeks, see a doctor.

It also cannot tell you whether your voice is pleasant or attractive. Those are listener judgements shaped by accent, context and familiarity, and no signal-processing tool measures them.

Where to go next