Sound: To LPC (autocorrelation)...

With this command you create a new LPC from every selected Sound, using the autocorrelation method.


You are advised not to use this command for formant analysis. For formant analysis, instead use Sound: To Formant (burg)..., which also works via LPC (linear predictive coding). This is because Sound: To Formant (burg)... lets you specify a maximum frequency, whereas the To LPC commands automatically use the Nyquist frequency as their maximum frequency. If you do use one of the To LPC commands for formant analysis, you may therefore want to downsample the sound first. For instance, if you want five formants below 5500 Hz but your Sound has a sampling frequency of 44100 Hz, you have to downsample the sound to 11000 Hz with the Sound: Resample... command. After that, you can use the To LPC commands, with a prediction order of 10 or 11.


Prediction order
the number of linear prediction coefficients, also called the number of poles. Choose this number at least twice as large as the number of spectral peaks that you want to detect.
Analysis window duration (s)
the effective duration of each analysis frame, in seconds.
Time step (s)
the time step between two consecutive analysis frames.
Pre-emphasis frequency (Hz)
a +6dB / octave filtering will be applied above this frequency. A pre-emphasis frequency of 48.47 Hz for a signal with a sampling frequency of 10 kHz approximately corresponds to a value of a = 0.97 for the filter yn = xn - a · xn-1. The relation between a and the pre-emphasis frequency is: a = exp (–2·π·preemphasisFrequency/samplingFrequency). If you do not want pre-emphasis, choose a frequency greater than the Nyquist frequency.


The autocorrelation algorithm is decribed in Markel & Gray (1976).

Links to this page

© David Weenink & Paul Boersma, April 7, 2004