I am doing research about emotion recognition from speech, by applying machine learning. Most papers are recommending using MFCC features. Therefore, I am currently trying to understand the underlying math and what MFCCs represent. I stumbled upon several new concepts like spectrum, cepstrum, etc. about which the papers assume I am already aware. As I have only a brief knowledge of sound processing, can someone guide me on how to approach this problem, what other concepts should I learn beforehand, and where to start?
If you don't know what a spectrum is, you're missing basically all of signal theory. So, get yourself an introduction to signals – there's many!
he first couple chapters of "Discrete-Time Signal Processing" by Alan V. Oppenheim and Ronald W. Schafer might pay off. Why?
- your university library has it. If they don't, they will buy it if you ask them to – it's really a standard book, no question.
- it's a good book, actually,
- it is the book that most of us agree on with respect to terminology, and
- it's super cheap used, if you want to buy it yourself.
And, again, you don't have to read all of it, but start at the beginning and take you somewhere until you understand spectra and filters.
Most papers are recommending using MFCC features
You're reading research papers – that might be a bit early for the moment, too.
You're operating in a mature research field, and that means there's textbooks that explain the basics quickly and in a didactically sensible way!
Any textbook on speech processing will take you further than reading papers without knowing the basics.
Hello maybe you’ll need these links to start and understand mel frequency cepstral coefficients.