mmmt-tools / Synthesis / LPC talkbox
on-device · nothing leaves this page

Linear predictive coding (LPC)

The model behind every 1980s talking toy. LPC fits an all-pole filter to each short frame of a sound — the filter that best predicts each sample from the last dozen — and treats what's left as the excitation. Separate the two and you can keep the filter (the moving vocal-tract shape, the words) but replace the excitation with a fresh buzz at whatever pitch you play: a talkbox. Swap the buzz for noise and the same words come out as a whisper — the source–filter split laid bare. The built-in source sweeps through vowels; drop your own voice to make it sing that. Analysis is offline; play it with the keyboard.

built-in voice (a-e-i-o-u)
waveform
spectrum (the tract shape)

Excitation

0.00
1.00×

Speed slides the words faster or slower without changing the pitch you play — time and pitch are independent in LPC.

Analysis

18
0.020
0.25 s

Higher order = more poles = sharper formants (and more of the pitch leaks into the filter).

hold a note (a s d f …) to voice the words at that pitch — it's monophonic, like a real talkbox