Week 9: Analysing and extracting meaning from audio

Task 1: Extracting features

Task 2: Computing and visualising histograms of features

Histograms of features computed from spectrogram

---

Histograms of features computed from MFCCs

Mel-frequency cepstral coefficients (MFCCs) capture a set of 20 features related to colour of a sound (timbre). Thus, the MFCC plots shown below can be used to discriminate between the songs based on the music instruments present.

---

Histograms of features computed from chromagrams

Chromagrams capture 12 features which are closely related to the twelve different pitch classes. Thus, the chromagrams shown below allow us to get a sense of the harmonic and melodic characteristics of the music studied.

---

Comparing findings from histograms

By looking at the histograms of the MFCC feature which commonly used for distinguishing between music instruments, we can clearly see that none of these tracks uses the same set of instruments since their histograms look a bit different. This makes sense as each track belongs to a different genre (rock, electronic, classical) and each of these genres uses a different set of instruments. The biggest outlier in terms of similarity is the song Air which is a piano song. Although, in my opinion, the tracks Malala (rock) and Indust (electronic) are not very similar when you actually listen to them, their MFCC histograms do not differ as much. This might be because besides other instruments, drums are fundamental to both genres.