Prsentation COST 277 Limerick

1
Non-Linear Speech Feature Extraction for Phoneme
Classification And Speaker Recognition
M. Chetouani, M. Faúndez-Zanuy (), B. Gas, J.L.
Zarader Laboratoire des Instruments et Systèmes
dIle-De-France (LISIF) Université Pierre Marie
Curie, PARIS, FRANCE () Escola Universitària
Politècnica de Mataró, BARCELONA, SPAIN
2
Outline

• Feature extraction in the recognition process
• Needs for speech feature extraction
• A non-linear model the Neural Predictive Coding
• Feature extraction for phoneme classification
• Feature extraction for speaker recognition
• Conclusions and future works

3
Speech Recognition Process

• Speech Feature Extraction Process
• Feature extraction is the first step of the
  recognition process.
• Feature extraction is usually computed by
  temporal methods like the Linear Predictive
  Coding (LPC) or frequential methods like the Mel
  Frequency Cepstral Coding (MFCC) or both methods
  like Perceptual Linear Coding (PLP).
• Limits
• Linear methods.
• No a priori class membership information
  (data-driven methods).

4
Feature extraction principle
Common part
Specific parts
5
Needs for Speech Feature Extraction (1)

• First, a non-linear modelization of the speech
  production model
• A solution Non-linear predictors (Volterra
  Filters, Neural networks)
• Our approach An extension of the Linear
  Predictive Coding (LPC) to the non-linear domain
  by neural networks.

6
Needs for Speech Feature Extraction (2)

• Problem generation of a great number of
  coefficients
• Solution
• First layer common to all the phonemes.
• One second layer specific to each phoneme.

7
The Neural Predictive Coding (NPC) (2)

• Principle
• The first weights w are common to all the phoneme
  classes.
• Each output layer is associated to one phoneme
  class.

8
Learning phase
9
Phoneme classification

• The objective is to extract from the speech
  signal phonetic features independently to the
  speakers.
• Common parts of the speech production model are
  modelized by the first layer.
• The specific parts are modelized by the second
  layers.

10
Discriminative Feature Extraction based on the
Minimum Classification Error (MCE) criterion

• The key idea is to provide discriminant
  constraints by classifiers way optimal
  constraints.
• Simultaneous training of both the feature
  extractor and the classifier
• Neural Predictive Coding.
• Prototyped-based classifier The Learning Vector
  Quantization.

11
Evaluation on phoneme classification

• Phonemes are extracted from the NTIMIT speech
  database
• DR1 and DR2 regions (without the SA sentences).
• 114 speakers for training and 37 for testing.
• Classification of confusable phonemes
• Vowels /ih/, /ey/, /eh/, /ae/,
• Voiced plosives /b/, /d/, /g/,
• Unvoiced plosives /p/, /t/, /k/
• Comparisons with traditional methods LPC, MFCC,
  PLP and a non-linear model NPC without explicit
  discriminant criterion.
• Classification by GMMs.
• Context independent classification
  (frame-by-frame).

12
Classification rates (frame-by-frame analysis)
13
Feature extraction for speaker recognition

• Objective
• speaker-dependent feature extraction.
• Speaker recognition process
• Feature extraction is currently carried out in a
  same way for all the speakers.
• Our approach
• A speaker model
• Feature extractor (NPC)
• Reference model.

14
A new initialization method for the NPC coding
phase

• Once the NPC model is parameterized, the coding
  phase consists in the estimation of the second
  layer weights.
• Like for all optimization processes, the
  initialization is important
• Traditionally, one uses random initialization
  with different constraints.
• We use the LPC analysis for linear initialization
  of the non-linear model
• Data-driven method for linear initialization of
  neural networks.

15
Speaker identification

• Modification of the traditional enrolment and
  test phases
• Enrolment phase
• NPC parameterization phase 12 seconds are used.
• Computation of reference models by using the
  whole sentence.
• Test phase
• The speech input is coded by each NPC model.
• The obtained features are compared by the
  associated reference models.

16
Evaluation on speaker identification

• 49 speakers from the Gaudi database
• Acquisition with a microphone connected to a PC.
• Vector dimension is set to 16
• One minute of read text is used for reference
  models training. From this minute, 12 seconds are
  used for the NPC parameterization.
• 5 sentences for testing (each sentence is about
  2-3 seconds).
• Comparisons with traditional coding methods LPC,
  MFCC, LPCC and PLP.
• Reference models computed by Covariance Matrices
  and the Arithmetic-Harmonic Sphericity (AHS) is
  used for comparisons.

17
Results
18
Conclusions

• The phoneme classification rates are improved by
  non-linear methods, in comparison with
  traditional methods like LPC, MFCC and PLP.
• The speaker identification rates are improved by
  non-linear methods with linear initialization in
  comparison with traditional methods like LPC,
  LPCC, MFCC and PLP.

19
Perspectives

• Phoneme classification
• Cooperation with different classifiers.
• Application to a large number of phonemes.
• Speaker recognition
• Explicit Discriminative Feature Extraction.
• Different applications identification,
  verification, tracking.

20

• Thank you for your attention