How do i use autocorrelation to find the fundamental frequency of a speech signal?

Question

Can anyone help me find the fundamental frequency (FF) of a speech signal with ta sampling frequency of 11025 Hz? Does the autocorrelation method work if so please explain. Do i have to divide the speech signal in to frames of 30 - 40 ms with 10 ms overlap to find the FF? And should o perform a Hann window on these frames?

Answer 1

Measure first the noise level of the detector.
Next take the digital data of the signal and subtract the noise level.

Break teh data stream into multiple groups (even overlapping) of uniform numbers of data, and figure out the Nyquist frequency (sampling limit).

For each group use a Hanning window to zero the ends to get a treated group.

Pass each treated group through an FFT and get a power spectrum.

Average the power spectrum of each group, plot and that should give you peaks at the fundamental frequencies.

Hope that helps

Answer 2

The Fourier transform might help, but not straight up.
It is really only good for deterministic signals.

A speech signal is more of a stochastic signal. The autocorrelation should work. Also, if you remember your properties, the autocorrelation function of a random process is the inverse Fourier transform of the power spectral density function. So that should help you get started.

Answer 3

Isn't it better to use fourier transform (or wavelett transform if you're interested in local FFs)?

Anyway, if there is a strong frequency component f then the autocorrelation will be highest (ideally +1) for lags that are a multiplum of 1/f, and -1 for lags that have the form (2n+1)/(2f). This is because cos(x+n/f)=cos(x) and cos(x+(2n+1)/(2f))=-cos x.