Extracting 12 Mel-frequency cepstral coefficients

[calebstrait]

Hi Yannick, I was hoping you might be willing to help me get past just one more issue:

I'm trying to extract 12 Mel-frequency cepstral coefficients (MFCCs) from a .wav file, and when I do the following:

sound = parselmouth.Sound(path)
mfcc_object = sound.to_mfcc(number_of_coefficients=12)
mfccs = mfcc_object.extract_features()
mfcc_arr = mfccs.as_array()
print(mfcc_arr.shape)

...the result on 3 different .wavs are

(4, 21)
(4, 2439)
(4, 114)

I'm expecting the resulting arrays to be 12 by (something), not 4 by (something). Where am I going wrong?

[YannickJadoul]

Hi @calebstrait. I thought I'd answered this as reply to the private email you had sent, but maybe something went wrong, there?

Anyway, I'll quote that reply here:

Your current problem is that Praat is annoyingly confusing, I think. The mfcc variable has type parselmouth.MFCC and already contains the MFCC coefficients. Calling extract_features() is doing some weird hidden action in Praat "To Matrix (features)..." (that used to be called "Extract features", I think?).
Anyhow, the comments in Praat's code say:

/* 1: cepstral difference function (d)
 * 2: spectral stability (dstab)
 * 3: spectral center of gravity (gs)
 * 4: stable internal duration
 */

So this is probably not what you want. Rather, you want to call mfcc.to_array() to get a numpy array containing the actual MFCCs. This should give a 13 by N matrix, (as the first feature contains the C0 value, related to the energy, and is not contained in the number_of_coefficients=12 argument, according to Praat).

Sorry that this is so confusing. Creating Parselmouth is always a thin line between trying to be clear from a Python point-of-view and trying to stay close to Praat :-/

By the way, there's also the Gitter chatroom for quick questions and answers, for things that are not per se an issue/bug.

[calebstrait]

Yes, thanks so much!! And apologies for missing your email response- looks like I need to update my spam rules.

[YannickJadoul]

No problem! Good luck with the rest of your project :)

[pranaswi23]

can u please give me the correct code of mfcc feature extraction because i am unable to understand your explanation.
Thanks in advance!!

[YannickJadoul]

can u please give me the correct code of mfcc feature extraction because i am unable to understand your explanation.
Thanks in advance!!

Asked (and answered) on Gitter: https://gitter.im/PraatParselmouth/Lobby?at=5efe111fc223cc536a4cf2a2

[pranaswi23]

Thanks for giving me the link of doccumentation but there is no explanation what the function is doing there.For example "to_mfcc" is there .Like it is taking no_of_coefficients as attribute but what is it doing ,no where mentioned.I am new to parselmouth and python.Please say how to know definition of a function in doccumentation.
I got wrong answer.Please tell me how to resolve .

from glob2 import glob
import librosa
import parselmouth

data_dir='D:\extraction'#folder path containing the .mp3 files
audio_file=glob(data_dir +'/*.mp3')
print(len(audio_file))
print(audio_file)

sound = parselmouth.Sound(audio_file[0])
print("sound",sound)
mfcc_object = sound.to_mfcc(number_of_coefficients = 12)
print("mfcc_object",mfcc_object)
mfccs = mfcc_object.extract_features(window_length=0.025)
print("mfccs",mfccs)
mfcc_arr=mfccs.as_array()
print("mfcc_arr",mfcc_arr)
print("mfcc_shape",mfcc_arr.shape)

i didn't understand your previous explanation how to get exact (12 mfcc coefficients) output.please explain.

[YannickJadoul]

@pranaswi23 That's what I answered above in this issue:

Calling extract_features() is doing some weird hidden action in Praat "To Matrix (features)..." (that used to be called "Extract features", I think?).

So this is probably not what you want. Rather, you want to call mfcc.to_array() to get a numpy array containing the actual MFCCs. This should give a 13 by N matrix, (as the first feature contains the C0 value, related to the energy, and is not contained in the number_of_coefficients=12 argument, according to Praat).

So extract_features is extracting different features, and you don't want to use that. Use to_array on the MFCC object.