Transcript PirknerMoodlogic

MoodLogic Metadata
MoodLogic
The promise and reality of digital music
The solution: Smart audio devices
Key technology: Creating music metadata
MoodLogic’s choice: Metadata generation with the help of end users
Mining of music string data (artist name, song name, etc)
Mining of individual song profiles
Mining of user collections and usage logs
© 2004 MoodLogic, Inc. – No reproduction or distribution without prior written permission.
- Confidential -
MoodLogic Metadata
The promise and reality of digital music
Promise:
• “Music at your fingertips …”
• “10,000 songs in your pocket”
• “The right music at the right time …”
Reality:
• 10,000 songs (many of them mislabeled or miscategorized)
only accessible via small screen real estate
• Issue: How to get the next hour of a great music experience?
- Confidential -
MoodLogic Metadata
The solution: Smart audio devices
What is a smart audio device?
• Content aware (not just portable hard disk but music device)
• Ability to create music experiences “on the fly”
• Adapt to the user preferences
Key: music metadata + inference technology
• Need for detailed descriptive data about individual songs
(genre, subgenre, mood, tempo, original release year,
instrumentation, etc)
• Playlisting algorithms: “Play an hour of smooth Jazz with
saxophone”, “Play songs similar to ‘Fight Music’ by D12”
- Confidential -
MoodLogic Metadata
Key technology: Creating music metadata
Table Of Contents Data
Classification Data (Metadata)
Basic TOC fields (Artist, Album, Song)
Used for:
Artist, Song Display
Tag fixing
There is no uniform database for music!
Detailed classification songs
Attributes (genre, mood, tempo, …)
Used for:
Browsing & Filtering, Playlist Creation
Recommendations
There is no perceptual database for music!
What are the options to generate metadata?
• DSP (Digital Signal Processing)
• Expert ratings/ submissions
• Community ratings/ submissions
- Confidential -
MoodLogic Metadata
MoodLogic’s choice: Metadata generation with the help of end users
Users listen to music
and fill out detailed
questionnaire
describing individual
songs
- Confidential -
MoodLogic Metadata
Mining of music string data (artist name, song name, etc)
• Mining 300 million submissions on (mis) spellings of artists, songs, albums
• Creating of a “canonical” artist space (e.g. making sure the same artist is spelled
the song the same way for all songs)
• Global database requires mining of music data in different languages and different
character sets
Dozens of different
spellings /
submissions for one
artist name for the
same song as well as
across songs by the
same artist.
Artist name
submissions for
song A:
Artist name
submissions for
song B:
Britney Spears
Britny Spears
Brittany Spears
Brittaney Spears
Britteney Spears
Britney Speas
Britney Spers
Britney Speares
…
Britny Spears
Brittany Spears
Goal: Uniform artist entities
- Confidential -
Britney Spears
Artist ID: 2435
MoodLogic Metadata
Mining of individual song profiles
• Mining > 1 billion individual song attribute ratings from end users (song model)
• Assessing quality of submissions (user model)
• Localization (different perceptions in different countries?)
• What are the salient attributes of a song?
Distribution of attribute “energy”
ratings for one song (rated by 18
people)
Goal: Determine song profile
- Confidential -
MoodLogic Metadata
Mining of user collections and usage logs
• Mining > 1 million user music collections (e.g. Finding like minded users)
• Building and evaluating quality of recommendation systems
• Determining the usefulness of product features
Distribution of music collections
for a million users
Song IDs for user A:
Song IDs for user B:
873124
243515
135123
646334
345321
246213
664343
621354
….
243515
135123
646334
345321
246213
664343
621354
542342
….
Goal: Recommend songs
- Confidential -
Song ID: 24543
MoodLogic Metadata
Questions? Interests? Suggestions?
Thanks!
- Confidential -