Music Information Retrieval

Presenting my findings in my paper Music Information Retrieval System, which was writen for the Noodle Study Tour to China and South Korea.

Transcript

1. 14 Octobe 1
2. 14 October 2013 2 THE PROBLEM
3. 14 October 2013 3 CURRENT SOLUTION: ASK NICELY
4. 14 October 2013 4 CURRENT SOLUTION: ASK NICELY
5. 14 October 2013 5 CURRENT SOLUTION: ASK NICELY •  Slow, especially for instrumental songs •  Ineffective for less popular music •  Requires user to identify distinct elements
6. •  What is an efficient search engine for identifying songs a user has heard? •  How do existing music search interfaces perform in identifying songs? •  Which techniques can be used to aid the music search engine? •  How do users perform when describing a song to identify it? 14 October 2013 6 RESEARCH QUESTIONS
7. •  Basic search: iTunes Store, Spotify: title, artist, album •  Lyrics search: Google, dedicated lyrics sites •  Query-by-Example: Shazam, Soundhound (singing the melody) •  Mood/occasion playlists: stereomood.com •  Tempo/Key search (for dj’s) •  Similarity suggestions (Last.fm, iTunes Genius, Pandora) 14 October 2013 7 STATE OF MUSIC SEARCH
8. •  Assisted narrow-down search •  Features both objective and subjective features •  Using automatic feature extraction algorithms 14 October 2013 8 THE IDEA A NEW WAY OF SEARCHING FOR MUSIC
9. Instrumental or with vocals? Instrumental Acoustical or Electronic production? Electronic Can you name a distinctive instrument? Accordion When do you think this song was released? Between 2000 and 2010 Are you looking for Samin – Heater? 14 October 2013 9 ILLUSTRATION HOW WOULD SUCH A SYSTEM WORK?
10. •  Technical •  Requires a database with millions of records with feature data •  Needs to be aggregated and generated •  Usability •  Can people actually use this? •  Plausibility •  Can users provide enough features to identify a distinct song? 14 October 2013 10 CHALLENGES & PLAUSIBILITY
11. •  Relatively new field •  Research in: •  Feature extraction (Tempo, Key, Genre) •  Instrument Recognition •  Mood and Emotional Classification •  Social Information Extraction •  Human Computation for Music Classification (Tagging games) •  Hit Song Science •  Applications are rare, and often proprietary. •  Resource expensive 14 October 2013 11 MUSIC DATA MINING STATE OF THE RESEARCH FIELD
12. •  User study •  Let a user pick a song from a set of 20 songs •  Then fill in a form of subjective and objective attributes •  Question the user on their affinity with music •  Let others guess the song on the basis of their answers •  Rinse and repeat •  Result: 17 people with different musical backgrounds filled in 94 song sheets 14 October 2013 12 USABILITY STUDY CAN USERS RECALL ENOUGH ATTRIBUTES TO IDENTIFY THE SONG?
13. •  70% of the filled in forms where guessed correctly •  84% of the forms filled in by people with high musical affinity •  People find it hard to classify Genre and Instrumentation •  People are good at remembering the tempo (BPM) of song, with a median deviation of just 7 BPM. •  Statistics and feedback hint that people get better at analyzing songs •  Positive feedback 14 October 2013 13 RESULTS OF USERSTUDY EARLY STATISTICS
14. 14 October 2013 14 AGREEMENT MOOD
15. 14 October 2013 15 AGREEMENT BEAT
16. 14 October 2013 16 AGREEMENT RHYTHM
17. •  Input of users gives a good chance of guessing the song correctly •  Requires a lot of resources •  System is technically challenging, but plausible 14 October 2013 17 CONCLUSION
18. •  Scalability needs to be researched •  Building this system is complicated and time expensive •  A user study on recalling songs after a longer time 14 October 2013 18 LOOKING FORWARD
19. 14 October 2013 19
20. 14 October 2013 20 STATE OF MUSIC SEARCH WHAT DO PEOPLE USE TO FIND MUSIC AND WHAT DO THEY SEARCH ON?
21. 14 October 2013 21 STATS
22. 14 October 2013 22 STATS