Singer similarity / identification

1
Singer similarity / identification

• Francois Thibault
• MUMT 614B
• McGill University

2
Introduction

• Relatively easy for humans to identify singing
  voice in various contexts
• Difficult to find time/environment invariant
  features for robust automatic identification
• Growing demand for such systems as Network
  databases keep expanding

3
Background (1)

• Significant research in speaker identification,
  systems perform poorly with singing voice
  (inadequate training)
• Singer identification research can draw much of
  automatic instrument recognition systems
• Artist / singer identification much harder than
  song identification (due to necessity of context
  invariant features)

4
Background (2)

• Often builds on speech / music discrimination
  systems
• Acoustical features heavily used to create
  N-dimensional Euclidean space loudness, pitch,
  brightness, bandwidth, harmonicity
• Often uses the same tools as style identification
  because each singer correspond to a micro style

5
Kim and Whitman overview

• Segmentation of vocal regions prior to singer
  identification algorithm
• Assumes singing regions display strong harmonic
  energy in voice frequency range
• Band-pass filter (200-2000 Hz)
• Inverse comb filter bank to detect harmonicity
• Identification classifier uses features based on
  LPC

6
K W features extraction

• Determine formant location and amplitude by a
  12-poles linear predictor using the
  autocorrelation method
• Augments low frequency resolution without
  increasing model order by warping the frequency
  representation with a function approximating the
  Bark scale

7
K W classification

• Uses Gaussian mixture model (GMM) to capture
  behavior of a class
• Parameters of Gaussians determined by Expectation
  Maximization (EM)
• Run PCA prior to EM (normalizes the data
  variance, good for EM)
• SVMs computes optimal hyperplane that can
  linearly separate classes

8
K W results

• Testbed contained more than 200 songs by 17 solo
  singers
• Half for training, half for testing
• Vocal segmentation inaccurate (55)
• Experimenting GMM and SVM for complete song and
  vocal parts only
• Overall results well short of human performance

9
K W Experimental results
10
Liu and Huang overview

• Singer classification of MP3 files
• First segment audio into phonemes
• Calculate feature vector and store phoneme
  feature vector with associated singer for
  training set
• Above feature vectors are used as discriminators
  for classification of unknown MP3 music objects

11
L H System Architecture
12
L H segmentation features

• Phoneme segmentation is derived from polyphase
  filter coefficients by obtaining a frame energy
  measurement

13
K W phoneme database

• Phonemes are separated by a minimum in FE

14
L H Phoneme features

• The phoneme features are obtained directly from
  the MDCT coefficients

15
L H classification (1)

• Compares phonemes features with those in the
  phoneme database
• Discriminating radius (Euclidean distance) is
  determines uniqueness of a phoneme
• Number of neighbors by same singer within the
  discriminating radius is called frequency (w)

16
L H classification (2)

• kNN classifier used to guess artist in unknown
  MP3 songs
• For efficiency, only uses the first N phonemes in
  unknown MP3
• Find the k closest neighbors in database and
  allow to vote if distance is within a threshold
• For each neighbor, give a weighted vote dependent
  on frequency, and distance

where w is frequency and
17
K W results

• 3 influencing factors
• Number of neighbors (N)
• Threshold for vote decision
• Number of singers in database

18
Other works

• Minnowmatch MIR engine including artist
  classification using NN and SVM (Whitman, Flake,
  Lawrence (NEC))
• Quest for ground truth in musical artist
  similarity determine accurate measure of
  similarity given subjective nature of artist
  classification (Ellis, Whitman, Berenzweig,
  Lawrence)