Finding a single voice in music Christine Smit April 26, 2007.

Finding a single voice in Finding a single voice in musicmusic

Christine SmitChristine Smit

April 26, 2007April 26, 2007

OutlineOutline

IntroductionIntroduction Classification Strategies: Classification Strategies:

Counting silent frequency binsCounting silent frequency bins Pitch cancellationPitch cancellation MFCCsMFCCs

Trading recall for precisionTrading recall for precision What worked and what didn’tWhat worked and what didn’t

IntroductionIntroduction

What am I doing?What am I doing?

What is a ‘single voice’?What is a ‘single voice’?

a a singlesingle note sounding at a note sounding at a timetime

Why do this?Why do this?

single voice finder + instrument single voice finder + instrument identifier identifier

==instrument sample libraryinstrument sample library

What are the data sets?What are the data sets?

training set: 10 1-minute samplestraining set: 10 1-minute samples test set: 10 1-minute test samplestest set: 10 1-minute test samples 25% single voice, 75% 25% single voice, 75%

multi-voice/silencemulti-voice/silence mixture of classical and folk musicmixture of classical and folk music

What characterizes a single What characterizes a single voice?voice?

non-solo solo non-solo

What characterizes a single What characterizes a single voice?voice?

What characterizes a single What characterizes a single voice?voice?

StrategiesStrategies

Strategy #1: Silence Strategy #1: Silence detectiondetection

findsilence

silent

HMM?

music

silencecounts

rawclassification

Nothing really workedNothing really worked

Strategy #2: Pitch Strategy #2: Pitch CancellationCancellation

music

filteredmusic

rawclassification

finalclassification

filterpitch

singlevoice?

HMM

Strategy #3: MFCCsStrategy #3: MFCCs

MFCC

GMM

HMM

music

13 features

likelihood

finalclassification

Trading recall for precisionTrading recall for precision

Quick reminderQuick reminder

PrecisionPrecision = out of the stuff we got, how = out of the stuff we got, how much of it was right? much of it was right?

Are google’s results relevant?

RecallRecall = out of all the right stuff, how = out of all the right stuff, how much did we get?much did we get?

If I asked google for the UN, did I get all the UN’s websites?

Precision is importantPrecision is important

If I have a large enough database, I If I have a large enough database, I can afford to have relatively low can afford to have relatively low recall. But I want high precision so recall. But I want high precision so what I do get is what I want.what I do get is what I want.

Strategy #2: Pitch Strategy #2: Pitch CancellationCancellation

music

filteredmusic

rawclassification

finalclassification

filterpitch

singlevoice?

HMM

Strategy #3: MFCCsStrategy #3: MFCCs

MFCC

GMM

HMM

music

13 features

likelihood

finalclassification

ResultsResults

Strategy #1: Silence Strategy #1: Silence detection detection (just for comparison)(just for comparison)

Strategy #2: Pitch Strategy #2: Pitch CancellationCancellation

Strategy #3: MFCCsStrategy #3: MFCCs

ConclusionConclusion

Silence detection really didn’t work Silence detection really didn’t work out.out.

MFCCs + GMM is really just as good MFCCs + GMM is really just as good as pitch cancellationas pitch cancellation

At 90% precision, I get about 25% At 90% precision, I get about 25% recall.recall.

AcknowledgementsAcknowledgements

Much thanks to Professor Ellis Much thanks to Professor Ellis for his assistance on this for his assistance on this

project.project.

Questions?Questions?