I'm trying to learn audio and voice/speech processing techniques out of interest. I intend to learn, experiment with algorithms and design, implementation of such systems, frameworks using programming languages like C/C++, Python, Matlab. Right now, to begin with, I'm only interested in audio or speech detection and not in speech recognition. I did some basic reading and research into this. Would like to understand a few basic concepts:

  1. How to detect the presence of human speech (voice frequency of male being between 85 Hz to 180 Hz, and that of female voice frequency between 165 Hz to 255 Hz but maximum range being 20 Hz to 20 kHz), differentiate them from other frequencies, identify noise etc?

  2. Being aware of the sampling theorem and the limitations of using lower sampling rates, how to know if any error was introduced during digital transmission or during reconstruction of the signal in analog domain from the digital domain (i.e., error detection) ?

  3. How to detect tones (produced digitally or externally recorded analog tones) and music notes? Do they follow similar techniques as speech? If no, how are they different?

I seek guidance, if not direct answers, on how to proceed with the learning the concepts and reading material. Any help in this regard will be grateful.

  • 1
    $\begingroup$ Your question is not in a format likely to attract answers. In general, it is preferred if you ask only one question, and if that question is very specific. Also, many of the subjects you want to learn on have multiple answers on this site, so it may be worth it to spend some time using the search engine. $\endgroup$ – MBaz Sep 27 '18 at 15:16
  • 1
    $\begingroup$ You might also want to research how pitch detection is different from frequency detection. This is a very important difference for speech and music. $\endgroup$ – hotpaw2 Sep 27 '18 at 15:33
  • 2
    $\begingroup$ I'm not sure your voice frequency numbers are correct, they look about right, but your units are off by a "k". $\endgroup$ – Cedron Dawg Sep 27 '18 at 15:43
  • $\begingroup$ @CedronDawg Oops. Thanks for pointing it out. Fixed it. $\endgroup$ – skrowten_hermit Sep 28 '18 at 3:43
  • 1
    $\begingroup$ @skrowten_hermit Given that you haven't got any answers yet, I'd suggest trying again, with one specific question per post. $\endgroup$ – MBaz Sep 28 '18 at 13:46

I'll give a few answers with the qualification that this isn't really my specialty area:

1) If you are using a DFT (FFT), then you should keep your sample frames relatively short. Speech utterances are of short duration and the DFT is best with steady tones across the entire frame.

2) Your sampling rate will likely be 44.1kHz (CD quality) or 8kHz, either of which is sufficient that the Nyquist frequency is nowhere close to the ranges you are talking about. In other words, this shouldn't be a concern.

3) Music tones tend to be of longer duration and a bit steadier. But long drawn out vowels fit the same pattern.

My advice would be get some software, like Audacity, that has a good spectrogram function and look at speech and music samples that way. Or write your own. Anyway, that is a good starting point to build up some familiarity with what various sounds and voices look like in the frequency domain.


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.