Crowd sourcing for tempo estimation

Improving perceptual tempo estimation with crowd-sourced annotations

Mark Levy, 26 October 2011

Tempo Estimation

Terminology: tempo = beats per minute = bpm

Tempo Estimation

Use crowd-sourcing: quantify influence of metrical ambiguity

on tempo perception improve evaluation improve algorithms

Perceived Tempo

Metrical ambiguity: listeners don’t agree about bpm typically in two camps perceived values differ by factor of 2 or 3

McKinney and Moelants: 24-40 subjects released experimental data

Perceived Tempo

Metrical ambiguity:

bpm bpm

liste

ners

liste

ners

McKinney and Moelants, 2004

Machine-Estimated Tempo

Also affected by metrical ambiguity: makes estimation difficult natural to see multiple bpm values estimated values often out by factor of 2 or 3

(“octave error”)

Crowd Sourcing

Web-based questionnaire: capture label choices capture bpm from mean tapping interval capture comparative judgements

Crowd Sourcing

Crowd Sourcing

Music: over 4000 songs 30-second clips• rock, country, pop, soul, funk and rnb, jazz,

latin, reggae, disco, rap, punk, electronic, trance, industrial, house, folk, ...

• recent releases back to 60s

Response

First week (reported/released): 4k tracks annotated by 2k listeners 20k labels and bpm estimates

To date: 6k tracks annotated by 27k listeners 200k labels and bpm estimates

Analysis: ambiguity

When people tap to a song at different bpm do they really disagree about whether it’s

slow or fast?

Investigation: inspect labels from people who tap differently quantify disagreement for ambiguous songs

Analysis: ambiguity

Subset of slow/fast songs: labelled by at least five listeners majority label “slow” or “fast”

Analysis: ambiguity

bpm vs speed label

all estimates for slow/fast songs

Analysis: ambiguity

bpm vs speed label

people can tap slowly to fast songs

all estimates for slow/fast songs

Analysis: ambiguity

Labels for fast songs from slow-tappers

Analysis: ambiguity

Quantify disagreement over labels: model conflict, extremity of tempo conflict coefficient

Ls, Lf, L: number of slow, fast, all labels for a song

L

LL

LL

LLC fs

fs

fs ),max(

),min(

Distribution of conflict coefficient C

Analysis: ambiguity

all songs with at least five labels

C > 0 means slow and fast

Analysis: ambiguity

Subset of metrically ambiguous songs: at least 30% of listeners tap at half/twice the

majority estimate

Compared to the rest: no significant difference in C

Evaluation metrics

MIREX: capture metrical ambiguity replicate human disagreement

Ambiguity considered unhelpful: automatic playlisting DJ tools, production tools jogging

Evaluation metrics

Application-oriented : compare with majority* human estimate

(*median in most popular bin) categorise machine estimates

same as humans twice as fast twice as slow three times as fast and so on unrelated to humans

Analysis: evaluation

Sources: BPM List (DJ kit, human-moderated)

Donny Brusca, 7th edition, 2011 EchoNest/MSD (closed-source algorithm)

maybe Jehan et al,? VAMP (open-source algorithm)

Davies and Landone, 2007-

Analysis: machine vs human

x2 same /2 unrelated other0%

10%

20%

30%

40%

50%

60%

70%

80%

BPM ListVAMPEchoNest

Analysis: controlled test

Controlled comparison: exploit experience from website A/B testing use this to improve algorithm iteratively

Result is independent of any quality metric

Analysis: controlled test

When visitor arrives at the page: choose a source S at random choose a bpm value at random choose two songs given that value by S display them together

Then ask which sounds faster!

Analysis: controlled test

Null Hypothesis: there will be presentation effects listeners will attend to subtle differencesbut these effects are independent of the source

of bpm estimates if the quality of the sources is the same

Analysis: controlled test

BPM List VAMP EchoNest0%

10%

20%

30%

40%

50%

60%

70%

80%

90%

100%

differentsame

Analysis: improving estimates

Adjust bpm based on class: imagine an accurate slow/fast classifier

Hockmann and Fujinaga, 2010 adjust as follows:

bpm:= bpm/2 if slow and bpm > 100 bpm:= bpm*2 if fast and bpm < 100 otherwise don’t adjust

simulation: accept majority human label

Analysis: adjusted vs human

x2 same /2 unrelated other0%

10%

20%

30%

40%

50%

60%

70%

80%

BPM ListVAMPEchoNest

Conclusions

Crowd sourcing: gather thousands of data points in a few days,

half a million over time humans agree over slow/fast labels, even

when they tap at different bpmImproving machine estimates: use controlled testing exploit a slow/fast classifier

Thanks!

mark@last.fm @gamboviol

http://mir-in-action.blogspot.comhttp://playground.last.fm/demo/speedohttp://users.last.fm/~mark/speedo.tgz

We are looking for interns/research fellows!