Timbre Recognition by Spectrum Analysisb.sturm/MAT201A/presentations/Fri/... · Timbre Recognition...

24
Timbre Recognition by Spectrum Analysis MAT201A 2008 Spring Project JungHo Ohn([email protected] ) Soo Hwan Park([email protected])

Transcript of Timbre Recognition by Spectrum Analysisb.sturm/MAT201A/presentations/Fri/... · Timbre Recognition...

  • Timbre Recognition by Spectrum Analysis

    MAT201A 2008 Spring Project

    JungHo Ohn([email protected])

    Soo Hwan Park([email protected])

    mailto:[email protected]

  • Agenda

    Motivation: Music Information RetrievalTimbre RecognitionAnalysis in time domain & freq. domain

    Features of Music Signal for Timbre RecognitionChallengesOur Approach Result

  • Music Information Retrieval (MIR)

    uplifting music & violin

  • Music Information Retrieval (MIR)

    Current Audio Search Engine- Semantic Information

    : Title/Tag information- MP3 Tag

    : Title, Artist, Album, Year,Comment, Genre Human Work

    - Query = Beatles & Let it be

    Future Audio Search Engine- Automatic Content-based Indexing- Cords, Notes, Timbre, Beat, Tempo, . - Query = Sad Song and Violin or

    Up-beat Blues, but Not Too Slow

  • Timbre/Instrument Recognition

    RecognitionA sound object currently being heard and its correspondence to something that has been heard in the pastMeasuring Similarity: melodic similarity, rhythmic and timbre similarities, genre and style similarities

    Timbrea quality of sound that distinguishes one music instrument from anotherImportant factor for Classifying musicSo far, no standard parameters

  • Analysis in time domain

    Correlation methodPitch informationSimilarity of a signal

    ( * )[ ] [ ] [ ]k

    f g n f k g n k

    =

    = +

    Test musical sound4A viola3C viola

  • Analysis in freq. domain

  • Spectrum AnalysisSpectrum AnalysisMusical sequence

    Triangular Window 2048

    4G_violin

    5A_flute

    4D_viola

  • Spectrum Analysis(cont.)

    Musical sequence

    4G_violin 5A_flute 4D_viola

    Spectrogram

  • Spectrum Analysis(cont.)

    Rect. Window0.5 sec

  • Spectrum Analysis(cont.)

    B1

    A1

    A2

    A3

    1 2 31 1 2 3B A A A = + +i i i

    1 11 21 31

    2 12 1

    3 2

    3

    1N N

    B A A AB AB

    B A

    =

    known known Unknown

    Least SquareEstimation

  • Features of Timbre in Frequency Domain (from PARK 2004)

    Harmonic Analysisbehavior of harmonics & locations of harmonics

    Inharmonicityerror between measured harmonics and their theoretical ideal harmonics

    Harmonic Expansion/CompressionHarmonic SlopeShimmer and Jitter

    Shimmer and jitter refer to short-time, subtle irregular variations in the amplitude envelope and frequency envelope of spectra respectively.

  • Harmonic Analysis (from PARK 2004)

  • Shimmer - short-time, subtle irregular variations in the amplitude envelope

  • Jitter - short-time, subtle irregular variations in the frequency envelope

  • Features of Timbre in Frequency Domain (from PARK 2004)

    Spectral EnvelopeSynchronicityTristiumulusSpectral CentroidSpectral IrregularitySpectral FluxLog Spectral SpreadRoll-offPhaseSpectral Flatness Measure

  • Features of Timbre in Time Domain (from PARK 2004)

    Amplitude Envelope: Attack, Steady-state, and DecayAttack: Rise Time and Slope Steady-state: the portion of a sound where the amplitude level is stable and a constant pitchDecay: portion of the envelope relates to the dying phase of a sound, ending with ultimate loss of energy

    Attack time (rise-time)Amplitude Modulation (Tremolo)

    typically ranges from 4~8 Hz and is usually found in the steady-state portion of a sound

    Temporal Centroidsignal descriptors for highly transient and percussive sounds

  • Attack, Decay, Sustain, Release(from Wikipedia)

  • Features of Timbre in Time Domain (from PARK 2004)

    PitchFrequency ModulationZero-Crossing Rate (ZCR)Linear Predictive Coding (LPC)Formants

  • Challenges

    Spectrum Analysis of Sample InstrumentsDefine our Features for Timbre RecognitionMake Decision Algorithm

    Correlation & Least Square Estimation

    Test with a Simple Synthesized Music

  • Our Approach

    STFT/FFTSTFT/FFT

    SampleData

    SampleData

    Feature Extraction(Harmonic,

    ..)

    Feature Extraction(Harmonic,

    ..)

    1. Training Phase

    Make References

    Make References

    2. Recognition Test

    STFT/FFTSTFT/FFT

    LoggingExtractedFeatures

    LoggingExtractedFeatures

    LeastSquare

    Estimation

    LeastSquare

    Estimation RESULT

  • Result Table

    Time4G

    violin

    2G_Boublebass

    4A_viola

    4D_viola

    5A_flute

    4E_violin

    5E_flute

    5D_flute

    2C_cello

    5F_flute

    0.5 0.95311 0.78844 0.02793 0.038758 0.093583 0.016083 0.27408 0.22503 0.1151 0.12256

    1 0.95311 0.78844 0.02793 0.038758 0.093583 0.016083 0.27408 0.22503 0.1151 0.12256

    1.5 0.3577 0.019156 2.7685 0.0969 1.1165 0.15592 0.41089 0.5101 0.1165 0.25907

    2 0.3577 0.019156 2.7685 0.0969 1.1165 0.15592 0.41089 0.5101 0.1165 0.25907

    2.5 0.95311 0.78844 0.02793 0.038758 0.093583 0.016083 0.27408 0.22503 0.1151 0.12256

    3 0.95311 0.78844 0.02793 0.038758 0.093583 0.016083 0.27408 0.22503 0.1151 0.12256

    3.5 0.4167 0.10644 0.2513 0.1953 0.5918 1.21157 1.63947 0.462 0.02712 0.12804

    4 0.95311 0.78844 0.02793 0.038758 0.093583 0.016083 0.27408 0.22503 0.1151 0.12256

    4.5 0.95311 0.78844 0.02793 0.038758 0.093583 0.016083 0.27408 0.22503 0.1151 0.12256

    5 0.4167 0.10644 0.2513 0.1953 0.5918 1.21157 1.63947 0.462 0.02712 0.12804

    5.5 0.4167 0.10644 0.2513 0.1953 0.5918 1.21157 1.63947 0.462 0.02712 0.12804

    6 0.34885 0.13649 0.227 1.46321 0.26759 0.22503 0.1534 1.60457 0.27162 0.17351

  • Result Sample Timbre Recognition

    Visualization with Processing(www.processing.org)

  • Thanks

    Any Questions?

    Timbre Recognition by Spectrum Analysis AgendaMusic Information Retrieval (MIR)Music Information Retrieval (MIR)Timbre/Instrument RecognitionAnalysis in time domainAnalysis in freq. domainSpectrum AnalysisFeatures of Timbre in Frequency Domain (from PARK 2004)Harmonic Analysis (from PARK 2004) Shimmer - short-time, subtle irregular variations in the amplitude envelopeJitter - short-time, subtle irregular variations in the frequency envelope Features of Timbre in Frequency Domain (from PARK 2004)Features of Timbre in Time Domain (from PARK 2004)Attack, Decay, Sustain, Release (from Wikipedia)Features of Timbre in Time Domain (from PARK 2004)ChallengesOur ApproachResult Table Result Sample Timbre RecognitionThanks