Voice Analysis Guides

Short, practical guides on how voice tests behave in the real world. Every chart on these pages comes from measurements we ran ourselves, mostly on real human recordings, with the method written next to it.

Most articles about voice scores repeat textbook ranges and stop there. These guides start from the opposite end: take real recordings, run them through the same analysis code the tools on this site use, and report what actually happened. When the numbers are small or the sample is narrow, the page says so.

The recordings are public-domain audiobook readings. The first three guides use three LibriVox readers: two healthy adult voices and one rough, low-quality recording. The pitch, resonance and progress guides use a larger set, 20 women and 20 men from the LibriSpeech corpus. An earlier calibration round used eleven readings, and where that data is cited it is labelled.

The guides

How these measurements are made

Each recording is converted to a mono signal and passed to the analysis file the site serves to every visitor, voice.js, running in Node.js instead of a browser tab. Nothing is tuned for the article. The same functions produce the score, age range and pitch you see after recording on the home page.

Where noise is added, it is generated with a fixed random seed, so rerunning the script gives identical numbers. The signal chain itself is documented step by step on How It Works, including the thresholds behind every sub-score.

Even forty readers is a small sample. These guides show how one analysis pipeline reacts to changes in take length, recording quality, pitch, resonance and signal type. They are not population statistics, and no page here should be read as a claim about how people in general sound.