2. Outline Introduction Music Information Retrieval Classification Process Steps Pitch Histograms Multiple Pitch Detection Algorithm Musical Genre Classification Implementation Future Work

3. Why do we classify? Increasing importance of digital music distribution Effectively navigating through large web-based music collections Structuring on-line music stores & radio stations Creating intelligent Internet music search engines and Peer-to-Peer systems Can be used in other type of analysis like similarity retrieval or summarization

4. Audio Classification Jazz Rock Classical Country Electronica Reggae World Folk New Age ? ? ? ? ? ? ? ? ? ? ? ? ?

5. Audio Classification (cont.)

7. Music Information Retrieval (MIR) The process of indexing and searching music collections. Symbolic MIR – Structured signals such as MIDI files are used. – Melodic information is typically utilized. Two different approaches: Query-by-melody (manual) and Query-by-humming Audio MIR – Arbitrary unstructured audio signals are used. – Timbral and rhythmic (beat) information is utilized.

8. What is MIDI? Musical Instrument Digital Interface A music definition language Communication protocol supports 128 different voices includes 16 channels

9. Classification Process Steps MIDI file Audio-from-MIDI fileArbitrary Audio file Pitch Histogram 4D Feature Vector (Pitch Content Feature Set) Multiple Pitch Detection Algorithm Labeled Feature Vectors used by Statistical Classifiers Histogram Construction Algorithm Timbral & Rhythmic Features Genre Classification Result by comparing the feature vectors

10. Pitch Histograms Unfolded Histogram –an array of 128 integer values (bins) indexed by MIDI note numbers –showing the frequency of occurrence of each note in a musical piece –contains information regarding the pitch range of the music Folded Histogram –All notes are transposed into a single octave and mapped to a circle of fifths –an array of 12 integer values –contains information regarding the pitch content of the music

11. Folded Pitch Histogram – Index Numbers Index Numbers 01234567891011 121314151617181920212223 242526272829303132333435 363738394041424344454647 484950515253545556575859 606162636465666768697071 727374757677787980818283 848586878889909192939495 96979899100101102103104105106107 108109110111112113114115116117118119 120121122123124125126127

12. Unfolded Pitch Histograms Fig.1 - Unfolded Pitch Histograms of 2 Jazz pieces (left) and 2 Irish songs (right).

13. Pitch Histogram features Four dimensional feature vector –PITCH-Fold –AMPL-Fold –PITCH-Unfold –DIST-Fold

14. Pitch Histogram Calculation For MIDI files: –The algorithm increments the corresponding note’s frequency counter while using linear traversal over all MIDI events in the file. –Normalization For arbitrary audio files: –Multiple Pitch Detection Algorithm

15. Multiple Pitch Detection Algorithm Fig.2 – Multiple Pitch Detection Flow Chart

16. Experiment Details Types of music contents: –symbolic (refers to MIDI) –audio-from-MIDI (generated using a synthesizer playing a MIDI file) –audio (digital audio files like mp3’s found on the web) Five musical genres are used: –Electronica, Classical, Jazz, Irish Folk and Rock Experiment Set: –A set of 100 musical pieces in MIDI format for each genre –A set of 100 audio-from-MIDI pieces for each genre –A set of 100 general audio files KNN(3) Classifier

17. Classification Results in MIDI Fig.3 – Classification accuracy comparison of random and MIDI

18. Classification Results in MIDI

19. Fig.4 – Pair-wise evaluation in MIDI

20. Classification Results in MIDI Fig.5 – Average classification accuracy as a function of the length of input MIDI data

21. Classification Results in Audio-from-MIDI Fig.6 - Classification accuracy comparison of random and Audio-from-MIDI

22. Classification Results in Audio-from-MIDI

23. Comparison of Classification Results Fig.7 – Classification accuracy comparison

24. Implementation  MARSYAS –MusicAl Research SYstem for Analysis and Synthesis –the software used for audio Pitch Histogram calculation and musical genre classification. –Three distinct modes of visualization: Standard Pitch Histogram plots 3D pitch-time surfaces Projection of the pitch-time surfaces onto a 2D bitmap

25. MARSYAS Visualization Fig.8 – Examples of grayscale pitch-time surfaces. Jazz (top) and Irish Folk music (bottom)

26. Summary Symbolic representation is more preferable in the sense of computing Pitch Information. This work can be viewed as an attempt to bridge the two distinct MIR approaches by using Pitch Histograms. Pitch Histograms do carry a certain amount of genre- identifying information. Multiple Pitch Detection Algorithm is not perfect, but it works by a certain degree.

27. Future Work Real-time running version of Pitch Histogram. –for better classification performance. –to conduct more detailed harmonic analysis such as figured bass extraction, tonality recognition, and chord detection. The features derived from Pitch Histograms might be applicable to the problem of content-based audio identification or audio fingerprinting. Alternative feature sets are needed. Query-based retrieval mechanism for audio music signals.

28. Thanks Cosku Turhan for the art work on my slides… 4 Non Blondes for their song, “What's Up?” :)