LPC Voice Generator — User Guide
Fixed LPC analysis–resynthesis: extracts pitch and a time-varying LPC envelope from the source, generates synthetic phonation, and filters it through the analyzed spectral envelope.
What this does
LPC Voice Generator analyzes the selected sound's pitch and time-varying LPC spectral envelope, generates a synthetic pitch-driven phonation source, and filters that source through the LPC model. The result is a synthetic voice-like rendering whose excitation is artificial but whose broad spectral shaping follows the analyzed source.
What is LPC?
Linear Predictive Coding / linear prediction models a short segment of speech as the output of an all-pole filter. In source–filter terms, the excitation supplies periodic or noisy energy, while the LPC filter approximates the time-varying resonant envelope associated with the vocal tract. LPC is especially effective on speech and other signals whose spectral envelope can be represented by a relatively small number of resonances.
Fixed processing
| Stage | Hard-coded setting |
|---|---|
| Pitch analysis | 75–600 Hz; Pitch object smoothed by 10 before conversion to PitchTier. |
| Excitation | PitchTier → phonation Sound at 44.1 kHz; adaptation 1, maximum period 0.05 s, open phase 0.7, collision phase 0.03, powers 3 and 4. |
| LPC analysis | Autocorrelation LPC, order 44, 25 ms analysis window, 5 ms frame step, 50 Hz pre-emphasis. |
| Synthesis | The synthetic phonation Sound is filtered through the time-varying LPC object. |
| Level | Praat Scale intensity: 70, i.e. 70 dB SPL in Praat's intensity convention, not 70 dBFS. |
The output object is renamed voice_synthesized and plays automatically.
Output and visualization
The Picture window compares the original and synthesized waveforms, plots the extracted 75–600 Hz pitch contour, and shows original and synthesized spectrograms up to 4 kHz. The summary reports LPC order 44, a 25 ms window, 50 Hz pre-emphasis, 44.1 kHz phonation and the 70 dB target intensity.
The script does not apply a final peak ceiling after Scale intensity: 70. The level control is therefore an intensity target rather than a clipping-protection stage.
Historical and technological context
The design belongs to the long source–filter / vocoder lineage. Homer Dudley's Bell Labs work in the 1930s separated speech into an excitation source and a spectral-shaping system, demonstrating that speech-like output could be reconstructed from a buzz or noise source passed through time-varying filters. Linear prediction later provided a compact mathematical model of that spectral envelope. Atal and Hanauer's 1971 analysis–synthesis system explicitly used predictor coefficients, pitch information and voiced/unvoiced excitation; Makhoul's 1975 review became a standard technical reference for LPC.
Compositionally, the important idea is that voice identity can be decomposed and recombined: pitch trajectory, excitation character and resonant envelope need not come from the same physical production mechanism. This script uses that separation in a deliberately synthetic way rather than attempting transparent speech reproduction.
Further reading
Dudley, H. (1939). “The Automatic Synthesis of Speech.” Proceedings of the National Academy of Sciences, 25(7), 377–383. DOI: 10.1073/pnas.25.7.377.
Atal, B. S., & Hanauer, S. L. (1971). “Speech Analysis and Synthesis by Linear Prediction of the Speech Wave.” Journal of the Acoustical Society of America, 50(2B), 637–655. DOI: 10.1121/1.1912679.
Makhoul, J. (1975). “Linear prediction: A tutorial review.” Proceedings of the IEEE, 63(4), 561–580. DOI: 10.1109/PROC.1975.9792.