-1
$\begingroup$

I am extremely new to signal analysis. And before posting I did a lot of reading on signal analysis, FFT and windowing. I am working on my thesis which involves comparison of speech signals lets say about 100 speech samples for a given sentence. I have the recordings and all the data. I have a few questions in order of what I think I should do.

  1. I need to be able to separate noise from the signal. What parameters should I use for that? If I have to use a window function then what sort of window should I apply for that? Hanning, Kaiser or Rectangular.

  2. What parameters should I look for to find out the similarities and the differences in the speech signals? Should I see the spectral densities, or the amplitudes and intensities?

Sorry for being so naive, I'm a real noob here. I hope you can help me out and bear with me patiently.

For software I am using PRAAT and for noise removal I think Audacity would be good. I used a Sony voice recorder for recording the speech samples.

$\endgroup$
  • $\begingroup$ Also I am a big failure at MATLAB.. takes me ages to get along with it.. that is why i chose the above softwares $\endgroup$ – Sulayemaan Jun 18 '14 at 17:24
  • $\begingroup$ 1. If I understand correctly, you are looking for noise removal, not for VAD algorithms? 2. Comparison of speech sounds based on their PSD is meaningless and it has too many dimensions - i can't imagine that, especially there are better features to do that. I suggest you to take a look into LPC coefficients, or at least ASE's. If you can, then perform the proper analysis with MFCC's - that's what these are invented for. Regarding MATLAB - it's the easiest programming language to learn (besides LabView) so it is not an excuse - MATLAB help manual is crucial. $\endgroup$ – jojek Jun 18 '14 at 20:18
  • $\begingroup$ 1. Yes i am looking for noise removal. In Audacity there is a noise removal tool and another tool labelled as Compressor. Do you think it will work? $\endgroup$ – Sulayemaan Jun 20 '14 at 5:13
  • $\begingroup$ 2. Also can you please let me know what does ASE stand for? For Analyzing MFCC would I be needing MATLAB? Thank you for your suggestions. I would definitely study about MFCC and LPC before asking further questions as I have no idea what they are ! $\endgroup$ – Sulayemaan Jun 20 '14 at 5:15
  • $\begingroup$ ASE - Audio Spectrum Envelope. It is a simplified version of spectrogram. You can find some answers on this site regarding MFCC's, but it is very easy to do in MATLAB thanks to many implementations laying around. As for noise removal algorithm in Audacity - you can do it, but I wouldn't expect great results. Definitely do not use compressor! It will make noise more prominent and decrease your SNR! $\endgroup$ – jojek Jun 20 '14 at 8:56
-1
$\begingroup$

Noise removal

You should use a Gaussian convolution filter.

Similarities in signal

Generally this is done by spectrum analysis - like a Fourier transform. get the DTFT of say every second or half-second (you will need to experiment with window size to get best results) and then match that to a database of Fourier transforms for your reference sounds. You'll probably want to pull out frequencies with the highest amplitudes and make a histogram, which you can query your database to find the closest histogram.

Sure, you could do parts of this with Audacity or other software, but if you want to learn more about DSP i suggest you use Matlab or Python+numpy/scipy to code the processing yourself! It will be a lot more flexible, maybe frustrating at times, but I highly recommend it.

$\endgroup$
  • $\begingroup$ Gaussian smoothing for the noise removal in speech signals? How it is better from other techniques meant to be used especially with speech? If you want to use it simply as the low-pass filter then I can't see why it is any better for 1D signals. In case of images they do have advantages. Any work/articles regarding approach you proposed and its accuracy? $\endgroup$ – jojek Jun 19 '14 at 0:42
  • $\begingroup$ Thank you for your response. Which software would i be requiring to apply a Gaussian convolution filter? MATLAB i guess? Regarding the spectrum analysis. I have quite a couple of windows available in Audacity. I presume using Rectangular is out of the question since it leaves a lot of leakage in the side bands. Do you think I should use a Gaussian window instead of Hanning? I'll post images of the spectrum analysis. (I don't understand what should be the criterion for best result.) $\endgroup$ – Sulayemaan Jun 20 '14 at 5:18

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.