poster research open day
TRANSCRIPT
Content-‐based Analysis of Last.fm Radio Sta6ons
1. Introduc6on
• Last.fm is an online music recommenda2on and streaming service that enables users to discover new music based on their previous listening experiences.
• The service allows for listening to radio sta2ons that are centered around a given ar2st or genre.
• Radio sta2ons are compiled based on the users’ listening history and social tags.
• This work explored the audio content of the sta2ons in order to gain insight into the data and improve the exis2ng service.
4. Analysis I – Visualiza6on of sta6on content
• Mul2dimensional scaling (MDS) of feature vectors of each sta2on.
→ all sta2ons showed a clearly iden2fiable centre
5. Analysis II – Homogeneity of sta6ons
• How far are the data vectors spread across the feature space?
→ rock-‐related sta2ons are most compact → techno-‐related sta2ons are the least homogeneous
8. Conclusion
• Analyses gave a deeper understanding about the organisa2on of the radio sta2ons and about the rela2ons of musical genres in general.
• Outlier detec2on procedure enables cleaning up radio sta2ons from poten2al unsuited audio tracks.
2. Overview
Audio content analysis:
radio sta2ons feature extrac2on analysis
Pop • track 1 • track 2 …
Jazz • track 1 • track 2 …
Blues • track 1 • track 2 …
7.2 3.5 1.6
5.1 8.6 0.7
track 1 track 2
feat. 1 feat. 2 feat. 3
…
3.0 4.1 2.9
1.6 9.7 1.4
track 1 track 2
feat. 1 feat. 2 feat. 3
…
6.3 7.8 0.1
5.7 3.9 1.1
track 1 track 2
feat. 1 feat. 2 feat. 3
…
Visualisa2on of content within each sta2on
Homogeneity es2ma2on
Outlier detec2on
Visualisa2on of sta2on rela2ons
3. Feature Extrac6on
Features from highest ranked algorithm of the MIREX 2009 “Audio Music Similarity and Retrieval” task:
• Timbre Features: • MFCCs (32 dimensions) • Spectral Contrast (32 dim.)
• Rhythm Features: • Onset Pacerns (125 dim.)
med
ian abs. dev.
6. Analysis III – Outliers
• Which tracks are different from most other tracks in a sta2on?
→ tracks with large amounts of silence → tracks with strong transients in generally non-‐percussive sta2ons → tracks with limited frequency response (e.g. vintage recordings) → speech tracks, a cappella tracks, tracks with less-‐suited content
7. Analysis IV – Sta6on rela6ons
• Mul2dimensional scaling of feature centers of all sta2ons
→ map of musical genres
Holger Kirchhoff [email protected]
Mark Sandler [email protected]