Song Intersection by Approximate Nearest Neighbours

1
Song Intersection by Approximate Nearest
Neighbours

• Michael Casey, Goldsmiths
• Malcolm Slaney, Yahoo! Inc.

2
Overview

• Large Databases Everywhere!
• 8B web pages
• 50M audio files on web
• 2M songs
• Find duplicates with shingles
• Text-based
• LSH - Randomized projections
• Results
• Best features
• 2018 song subset

3
The Need for Normalization

• Recommendations
• Apply one songs rating to another
• gt Better matches
• Playlists
• Find matches to user requests
• Remove adult/child music
• Search results
• Dont show duplicates

4
Specificity Spectrum
Cover songs
Remixes
Fingerprinting
Genre
Look for specific exact matches
Bag of Features model
Our work (nearestneighbor)
5
Remixes of One Title
6
Remix Examples
Abba Gimme Gimme
Madonna Hung Up
Tracy Young Remix of Hung Up
Tracy Young Remix 2 of Hung Up
7
How Remix Recognition Works

• Algorithm
• Matched filter best (ICASSP2005 result)
• Nearest neighbor in 3601200D space
• Ill posed?
• Efficient implementation
• Audio shingles
• Like web-duplicate search
• Locality-sensitive hashing
• Probabilistic guarantee

8
Audio Processing
9
Remix Distance
Matched filter (implemented as nearest neighbor)
N-best matches
10
Choosing r0
11
Hashing

• Types of hashes
• String put casey vs cased in different bins
• Locality sensitive find nearest neighbors
• High-dimensional and probabilistic
• Two Nearest Neighbor implementations
• Pair-wise distance computation
• 1,000,000,000,000 comparisons in 2M song database
• Hash bucket collisions
• 1,000,000,000 hash projections

12
Random Projections

• Random projections estimate distance
• Multiple projections improve estimate

13
Locality Sensitive Hashing
Distant Vector

• Hash function is a random projection
• No pair-wise computation
• Collisions are nearest neighbors

Distant Vector
14
Remix Nearest Neighbour Algorithm 1

1. Extract database audio shingles
2. Eliminate shingles lt songs mean power
3. Compute remix distance for all pairs
4. Choose pairs with remix distance lt r0

15
Remix Nearest Neighbour Algorithm Revisited

1. Extract database audio shingles
2. Eliminate shingles lt songs mean power
3. Hash remaining shingles, bin widthr0
4. Collisions are near neighbour shingles

16
Method

• Choose 20 Query Songs
• Each has 3-10 Remixes
• 306 Madonna Songs
• 2018 MadonnaMiles

17
Results
18
Conclusions

• Remixes are hard, but well-posed
• Brute force distances too expensive
• LSH is 1-2 orders of magnitude faster
• LSH Remix Recognition is Accurate

19
Conclusions

• Remixes are hard, but well-posed
• Brute force distances too expensive
• LSH is 1-2 orders of magnitude faster
• LSH Remix Recognition is Accurate

20
Conclusions

• Remixes are hard, but well-posed
• Brute force distances too expensive
• LSH is 1-2 orders of magnitude faster
• LSH Remix Recognition is Accurate

21
Conclusions

• Remixes are hard, but well-posed
• Brute force distances too expensive
• LSH is 1-2 orders of magnitude faster
• LSH Remix Recognition is Accurate