Applying a window function to a speech signal

I'm currently developing a speaker recognition system. I have two main problems:

1. What is the most suitable windowing function to apply to a speech signal prior to the STFT?

2. Is it necessary to apply more than one windowing function?

• The "most suitable" window depends on exactly what kind of processing you need or plan to do with the data after windowing (and after the FFT). One window type may have better frequency response in certain ranges, another better phase response, yet another might help with more accurate amplitude estimation. & etc. Often a ( what every body else is using ) generic widow is used as a compromise. – hotpaw2 Dec 28 '16 at 4:50
• @hotpaw2 I am planning to obtain the spectrogram after performing fft. This is to extract features of the voice sample. Will a Hamming window be sufficient? – mgw2016 Dec 28 '16 at 9:29

For speech signals usually hanning or hamming window with 50 percent overlap is used.

What is the most suitable windowing function to apply to a speech signal prior to the STFT?

hanning or hamming window with 50 percent overlap is used generally.

Is it necessary to apply more than one windowing function?

Not necessary i would say.

• there is no "Mr. Hanning" nor "Dr. Hanning". however there is a Julius von Hann with a window named after him. – robert bristow-johnson Dec 27 '16 at 20:50
• @robertbristow-johnson Thanks !! , so shall I name the windows with lowercase letters ? – Arpit Jain Dec 28 '16 at 5:00
• Usually, this type of window is called "Hann window" or, however less often, "von-Hann window". – applesoup Dec 28 '16 at 8:13
• @arpitjain Yes, I'm going to use the Mel scale after performing FFT to the voice sample. – mgw2016 Dec 28 '16 at 9:40

I don't really understand what "apply a suitable windowing function" "prior to the STFT" means. Most STFT implementation incorporate apodizing windows. And the best is always related of some cost function, objective or processing.

Note that there are people called Mr Hanning, but they are not (AFAIK) at the origin of the raised-cosine or von Hann window (Julius von Hann, Austrian meteorologist). You can find a mention at page 97 from the book The measurement of power spectra from the point of view of communications engineering (archive.org), 1958, by Blackman and Tuckey. Noteworthy, they spell hamming and hanning in small letters.

Hamming or Hann windows are very standard. $50\%$ or $75\%$ overlap are common. In the recent works I know of, especially for blind source separation, people often use at least two window sizes: a shorter one for transients, and a longer one.

This is a first step toward using different windows depending on the scale of observation. Some have used a multiscale context, such as in the
ERBlet non-stationary Gabor filterbank, ERB for Equivalent Rectanguar Bandwidth, with a constant-Q transform:

Some have been using asymmetric windows, better suited to the skewed shape of speech onsets (see eg Using asymmetric windows in automatic speech recognition).

The choice of the best windows, to me, even for speech, is not clearly settled.