I've been trying to wrap my head around MFCC for a while now, and I have some questions.
My understanding is that MFCC is basically windowing, framing, and Fourier Transforming (STFT) the sound signal, and applying Mel filter bank to it, and log to it and doing Discrete Cosine Transform to get the MFCCs. I think these coefficients are supposed to help learn stuff about the sound, but I'm unclear on how it does that. They help us learn about the formants or something, but I'd like a more detailed answer. I have several questions. I guess questions 2 and 3 are sort of the same.
- I know there will be orders of the coefficient, so what do lower coefficients tell us? What do higher coefficients tell us?
- Why do we even use MFCC in the first place? What does it provide that a normal spectrogram couldn't?
- How do you even use a cepstrum? I know the x-axis is quefrency, and I saw somewhere that used (sampling rate)/quefrency to find the frequency, but what does this tell us?
- What does the Discrete Cosine Transform do? Why do we need it?
Please be detailed in answering me, because I've been to many websites trying to learn, but I'm still confused. Thank you.