Music Mood Detection (Lyrics based Approach)

3,987 views

Published on

A presentation on Music Mood detection techniques including information retrieval through audio features and lyrical features.

0 Comments
7 Likes
Statistics
Notes
  • Be the first to comment

No Downloads
Views
Total views
3,987
On SlideShare
0
From Embeds
0
Number of Embeds
8
Actions
Shares
0
Downloads
218
Comments
0
Likes
7
Embeds 0
No embeds

No notes for slide

Music Mood Detection (Lyrics based Approach)

  1. 1. Presented by: Akhil H. Panchal T.E. Computer Guided by: Prof. Mrs. Tiple Computer Dept. 1
  2. 2. CONTENTS  Mood vs. Emotion  Why MMD?  Mood Models  How MMD?  Audio Features  Hierarchical MMD algorithm  Lyrics Features  A Lyrics based approach to MMD  Applications  Limitations 2
  3. 3. EMOTION! • Reactions to an event or a stimulus that lasts for a short period of time. • Important concern for Music psychologists. 3
  4. 4. MOOD! • A generalized form of your emotional feelings that last for a longer period of time. • Less intense. • Important concern for MIR researchers! 4
  5. 5. WHY MMD?  Need for sorting the ever increasing Music Database according to our choice(mostly being “Mood”).  Time consuming for Listeners to manually select songs suiting a particular mood or occasion.  Huge variety of our Music ranging from various Albums/Artists/Composers which is heavily influenced by mood. 5
  6. 6. MOOD MODELS!  A way to classify various moods so that each mood can be identified distinctively. Mood Models Categorical Dimensional 6
  7. 7. HEVNER’S MODEL 7
  8. 8. RUSSELL’S MODEL 8
  9. 9. THAYER’S MODEL 9
  10. 10. NAVRAS : INDIAN CLASSICAL MODEL 10
  11. 11. HOW? Music Mood can be detected by 2 main techniques. 11
  12. 12. AUDIO FEATURES 2-tier taxonomy of Music Features: Low Level Time Signature Tempo(BPM) Timbral Temporal Mid & High level Pitch Rhythm Harmonies 12
  13. 13. AUDIO FEATURES  Low-level features not closely related to the properties perceived by ‘listeners’.  Mid-level features derived from low-level features help in extracting properties of Music closely perceived by ‘listeners’ as Mood. 14
  14. 14. LIST OF FEATURES Spectral Centroid Spectral Flux Mel- frequency Coefficients Roll-off point Zero- crossings Beat Histogram 15
  15. 15. Conversion of Hertz into Mel scale: 16 C=1127.01048
  16. 16. HEIRARCHICAL MUSIC MOOD DETECTION ALGORITHM 1. Start. 2. Convert Music clip into uniform format. 3. Divide Music clip into plurality of frames. 4. Extract Audio features: Spectral features, Beat histogram, Mel-frequency coefficients. 5. Calculate average frame intensities. 19  Based on Thayer‟s Mood Model  Used for classifying a music clip into either of the 4 categories: G1(Exuberance, Anxious),G2(Contentment & depression).  Algorithm:
  17. 17. HEIRARCHICAL MUSIC MOOD DETECTION ALGORITHM 6. Classify Music clip into a mood group based on intensity feature. a) Determine probabilities of 1st n 2nd group based on intensity. b) If P(G1)>P(G2) then select G1. Else select G2. 7. Classify Music clip into exact Music mood based on timbral & rhythm features. a) Determine probabilities of 1st n 2nd group based on intensity. b) If P(M1)>P(M2) then select M1 Else select M2. 20
  18. 18. LYRICS FEATURES Text Stylistic N-gram content words POS(Part of Speech) ANEW & WordNet General Enquirer LYRICS BASED APPROACH 21
  19. 19. TEXT STYLISTIC FEATURES  Include text statistics such as:  No. of unique words  No. of unique lines  No. of repeated lines/words  Words per minute  Special punctuation marks(!) &  Interjection words (e.g.: „Hey‟, „Oh‟) 22
  20. 20. PART OF SPEECH (POS) FEATURES  Grammatical tagging of words according to their definition and the textual context they seem in.  E.g.: Time flies like an arrow. (noun) (verb)(prep.)(art.) (noun) 23
  21. 21. N-GRAM CONTENT WORDS  Combination of unigrams, bigrams & trigrams of content words.  Help in detecting emotion. Happy Romantic Aggressive Hopeful Heaven With you I‟ve never If you All around Love Kill Dreams 24
  22. 22. ANEW & WordNet  ANEW has 1034 English words with scores in 3 dimensions:  Arousal  Valence  Dominance  Extended by adding synonyms from WordNet & WordNet-affect. 25
  23. 23. LYRICS BASED MOOD DETECTION SYSTEM  The lyrics of the song are given as input in textual form.  Lyrics pre-processing is performed.  Intro, Verses, Chorus are detected at this stage.  Instructions like „repeat chorus‟ are replaced by the actual lyrics.  Spelling errors are corrected. 26
  24. 24. LYRICS BASED MOOD DETECTION SYSTEM  Lyrical features mentioned are extracted (with help of ANEW, WordNet)  The song is tagged with various moods with varying probabilities.  The mood tagged with maximum probability is selected as the mood of the music clip. 27
  25. 25. CURRENT MMD PLATFORMS  Stereomood.com  Musicovery.com  Mymusicsource.com  Last.fm  Youlicense.com  Crayonroom.com  Googlemusic.com (China) 28
  26. 26. 29
  27. 27. APPLICATIONS  Shop owners seeking music to attract certain clients.  Sorting the music that we have according to a certain mood or occasion.  Ad films requiring a highly memorable & positive emotion invoking music for their products. 30
  28. 28. APPLICATIONS  A Disk Jockey seeks Music having the same beat & a similar mood as the current song.  In games, to invoke moods such as excitement, danger, fear, victory & happiness.  A call center asking the callers to hold, need happy music pieces. 31
  29. 29. LIMITATIONS  Precision issues in case of metaphors.  Mood from some Music pieces can be subjective.  Mood perceived highly dependent on cultural background.  Conversion to standard format leads to loss of certain features. 32
  30. 30. ♫Q & A♫ THANK YOU! 33

×