Sung consonants from a physical model of the voice

Every sound on this page is computed by a physical model of the human voice: vocal folds simulated as vibrating tissue driven by air pressure from the lungs, and a simulated vocal tract (throat, mouth and nose) whose shape moves for each sound. Nothing is recorded or sampled; you hear the computed sound pressure leaving the lips and nostrils.

The model sings a major scale, do re mi fa sol la ti do', from E3 (165 Hz) to E4 (330 Hz), each note one consonant followed by the vowel /ɑ/ (as in father); the first row is the vowel alone. Two voices sing it: Voice X, the model's main voice (vocal tract 15.8 cm, between typical adult female and male lengths), and a female voice (vocal tract 14.7 cm, with a wider gap at the back of the vocal folds). Both use the same vocal-fold model; singing the same notes puts the female voice in the low part of a woman's range.

What to listen for: whether you can tell which consonant is sung, how cleanly it joins the vowel, and whether each note lands on its pitch and holds steady. Under each clip is the result of an automatic check made on the vowel part of each note (it does not judge the consonant): a note passes when its pitch is within 50 cents (half a semitone) of the target, it does not drop to half the pitch, and its cycles repeat regularly. Each clip is a single take with the same random seed; none was picked as a best take.

Each clip is played in two radiation settings, side by side. Radiation is how sound leaving the mouth spreads into the air. The standard radiation follows the physics of a small opening: the higher the frequency, the more strongly it radiates, about 6 dB more per octave up to several kilohertz. The darker radiation is an older setting of the model that stops this rise at 400 Hz; it sounds smoother but dulls the high frequencies, including the hiss of consonants. The clips are matched in RMS level, so the two settings differ mainly in brightness.

For singing, the folds carry extra internal damping, an older setting of the model. Without it, these voices break on the top notes, ti and do', where the folds fall into a pattern that repeats every second cycle (heard as a rough voice an octave lower), and after many consonants they also slip at fa or sol. With it, every note of these clips passes the check with both radiation settings (368 of 368 each). Speech does not use it: even half this damping stops the folds from closing in speech and changes the voice.