By Tania Alwan
When people say "vocal cords," they mean the vocal folds: two bands of tissue in the larynx that open for breathing and come together to vibrate for speech and singing. They are not guitar strings, and they do not work by being plucked. This post walks through the standard acoustic model of how they make sound, then builds a tiny vowel synthesizer in Python so you can hear the difference between pitch and vowel.
What the vocal cords actually do
The current standard explanation of how the folds vibrate is the myoelastic-aerodynamic theory. In plain language:
- Air pressure from the lungs builds up below the nearly closed folds
- The folds are blown apart for a moment
- They snap back because of their elasticity and the drop in pressure as air rushes through
- That opening and closing happens many times a second and chops the airflow into a pulse train
The rate of those pulses is the fundamental frequency, f0. That is pitch. A lower speaking voice might have an f0 around 100 to 120 Hz; a higher one might sit closer to 200 Hz. There is a lot of overlap between voices, and singing moves well outside a person's speaking range.
Source and filter
The classic source-filter theory of speech production (Fant, 1960) splits the voice into two parts:
- The source: the glottal pulse train from the vocal folds. It has energy at f0 and at many harmonics (2f0, 3f0, and so on)
- The filter: the vocal tract (throat, mouth, lips, sometimes the nose), which resonates at certain frequencies called formants
Change the fold vibration rate, and the pitch changes. Change the shape of the tongue and lips, and the formants move, which is how /ɑ/ ("ah") sounds different from /i/ ("ee") on the same note. That is why you can sing the same pitch on many vowels.
The linear source-filter model is a simplification: in real singing the tract and the source can interact. It is still the most useful first model, and it is the one we will code.
A short vowel synthesizer in Python
This script builds a simple glottal pulse train, then sends it through three resonators tuned to a widely cited textbook approximation of an adult male /ɑ/ vowel: formants near 730, 1090 and 2440 Hz (Peterson and Barney, 1952). A first-difference at the end stands in for lip radiation, which tilts the spectrum brighter.
Install with pip install numpy scipy.
import numpy as np
from scipy.io.wavfile import write
from scipy.signal import lfilter
sr = 22050
duration = 1.0
f0 = 120 # Hz, the "pitch"
n = int(sr * duration)
source = np.zeros(n)
# Simple glottal-ish pulse: a short decaying burst once per period
period = int(sr / f0)
pulse_len = max(5, period // 5)
t_pulse = np.arange(pulse_len)
pulse = t_pulse * np.exp(-t_pulse / (pulse_len / 5))
pulse /= pulse.max()
for i in range(0, n - pulse_len, period):
source[i : i + pulse_len] += pulse
def formant(x, freq, bw, sr):
"""Two-pole resonator at freq Hz with bandwidth bw Hz."""
r = np.exp(-np.pi * bw / sr)
theta = 2 * np.pi * freq / sr
b = [1.0]
a = [1.0, -2 * r * np.cos(theta), r * r]
return lfilter(b, a, x)
# Approximate /ɑ/ formants (Hz) and bandwidths (Hz)
y = source
for freq, bw in [(730, 90), (1090, 110), (2440, 140)]:
y = formant(y, freq, bw, sr)
y = np.diff(y, prepend=0) # lip radiation
y /= np.max(np.abs(y))
write("vowel_a.wav", sr, (y * 0.9 * 32767).astype(np.int16))
Play vowel_a.wav. Then change f0 to 220 and run it again: same vowel, higher pitch. Or keep f0 = 120 and switch the formants to a typical /i/ approximation, around 270, 2290 and 3010 Hz (also from that same classic dataset, for an adult male /i/), and you should hear the vowel change while the pitch stays put.
This is a teaching model, not a singing-voice synthesizer. A real voice has changing f0 (vibrato, intonation), moving formants, breath noise and a much more detailed glottal pulse.
Why this matters if you work with audio
- Pitch trackers estimate f0, the source rate. They can be confused by strong formants, which is one reason a bright vowel or a mix can fool a tuner.
- Vowel identity lives in formants, not in f0. That is why a spectrogram of a sung /ɑ/ and /i/ looks different even on the same note.
- Voice conversion and AI singing models have to deal with both layers: keep or retarget the source (pitch, timing, expression) and change the filter (the timbre of another voice). Platforms such as Voicestars expose that idea as a catalogue of singer-style voice models; open-source voice-conversion projects let you run a similar process locally.
If you want to go one step further, plot a spectrogram of vowel_a.wav (librosa or scipy) and look for the three dark formant bands. They should stay put when you change f0, and jump when you change the resonator frequencies. That picture is the source-filter model, drawn.
Takeaways
- "Vocal cords" are vocal folds that periodically interrupt airflow; the rate of interruption is pitch
- The vocal tract filters that source; its resonances (formants) shape vowels and timbre
- A few lines of Python are enough to hear those two controls separately
Disclosure: this article was written in partnership with Voicestars.
Top comments (0)