question about feature extraction

[AliD201]

i see in feature extraction after this piece of code

audio_data, sample_rate = librosa.load(file_name) stft = np.abs(librosa.stft(audio_data))

that you some time use stf and some times use audio_data
like

mfcc = np.mean(librosa.feature.mfcc(y=audio_data, sr=sample_rate, n_mfcc=40).T,axis=0)
    features.extend(mfcc) # 40 = 40  

chroma = np.mean(librosa.feature.chroma_stft(S=stft, sr=sample_rate).T,axis=0)
features.extend(chroma) # 12 = 52`

why is that ? what is the difference ? would like you to elaborate more on this. :)

Thank you !