Record or upload • Track formants • Replace with sinusoids
Sine-wave speech replaces the first three formants of an utterance with time-varying sinusoids that follow their frequency and amplitude. Naive listeners usually hear whistles; once told (or shown the original) it is speech, most can suddenly understand it — a classic demonstration of top-down, prior-driven perception.
What is a formant? When you speak, your vocal folds produce a buzz rich in harmonics, and your vocal tract (throat, mouth, lips) acts as a set of resonators that amplify some frequency bands. These resonance peaks are the formants: F1 is the lowest, F2 the next, and so on. Their positions depend on the shape of your tract — mostly tongue height (F1) and tongue frontness (F2) — so moving formants are what distinguish one vowel from another, and their rapid transitions carry much of the consonant information. They show up as the dark horizontal bands in the spectrogram above.
This tool was coded by Claude (Anthropic's AI) as a JavaScript port of Chris Darwin's original Praat SWS script. It uses Burg LPC formant analysis (5 formants, 25 ms Gaussian window, 10 ms step), Viterbi tracking of 3 formants against talker-dependent references, amplitudes read off a wide-band (3 ms) spectrogram at each formant frequency and low-passed, then phase-integrated sinusoids at the original sample rate.