Vowel Space Live

A real-time vowel chart. Speak or sing, and the first two formants of your voice are traced on the F1–F2 plane as you go. Calibrate it to your own voice, compare yourself with the vowels of a language, practise hitting a target vowel, and record your track to replay or export.

F1—

F2—

Pitch (f0)—

Level · SNR —

Microphone

Voice

Sustained vowels and slow glides work best. Consonants, breaths and noise are skipped. Start with a second of silence so the page can learn the room noise. A clear, moderately loud voice works best; the SNR readout turns amber below 18 dB and red below 12 dB, where soft vowels (especially [u] and [i], whose F2 is weak) may drop out.

1 · Calibrate to your voice

Every speaker's vowel space is a little different. Hold three vowels for about two seconds each, and the chart adapts to you. A red cross shows what is being measured; if a vowel looks wrong, press “Redo this vowel”.

Hold

Not calibrated: using the voice type above.

2 · Trail

Fade after

keep the whole trace

3 · Reference vowels

Language

Target vowel

Pick a target and try to keep the dot inside its circle.

4 · Record

The chart. The horizontal axis is the second formant F2 (front vowels on the left), and the vertical axis is the first formant F1 (close vowels at the top), as in a phonetician's vowel chart. Before calibration the axes show your formants in Hz, scaled by the voice type. After calibration they show normalized values: your vowels are mapped onto an adult-male reference space, so you can compare yourself with the reference vowels.

Calibration. Your “ee”, “ah” and “oo” (each measured from the spectrum averaged over the hold) define a mapping, affine in log-frequency, from your vowel space onto the corner vowels [i], [a], [u] of the selected language, so your corner vowels land exactly on that language's. If [u] cannot be measured, [i] and [a] alone scale each axis separately. Working in log-frequency matches the main source of speaker differences: vocal-tract length scales all formants by roughly the same factor. The calibration is remembered in this browser.

Tracking. The microphone signal is low-pass filtered and resampled to 11 kHz. In each 30 ms window a 14th-order LPC fit gives resonance candidates, and an autocorrelation test checks that the window is voiced. A tracker picks F1 and F2 by sharpness, rank and continuity, rejects pairs outside the vowel space, and accepts a sudden jump only when several consecutive frames agree. The background noise spectrum is learned during pauses; a resonance that does not rise clearly above it is not accepted as a formant, so a voice that is too soft shows a gap rather than a wrong point.

Reference vowels. Approximate averages for adult male speakers, from published acoustic studies; real speakers vary a lot, and dialects differ. Treat the circles as neighbourhoods, not exact targets. High voices (above about 250 Hz) make formants hard to measure for any method, because the harmonics sample the spectrum too sparsely.

Recording. Records your voice together with the tracked values (about 60 per second); both stay in your browser. Replay plays the audio while the trace redraws in step with it, and the player's position bar lets you jump to any moment. The WAV (mono, 16-bit, at the microphone's sample rate) can be opened in Praat; the CSV has time, voicing, f0, raw F1 and F2 in Hz, the F1 and F2 shown on the chart, and in its header the track time at which the audio starts.