ON REALTIME MEANANDVARIANCE NORMALIZATION OF SPEECH RECOGNITION FEATURES

1
ON REAL-TIME MEAN-AND-VARIANCE NORMALIZATION OF
SPEECH RECOGNITION FEATURES

• Pere Pujol, Duan Macho, and Climent
  NadeuNational ICT
• TALP Research CenterUniversitat Politècnica de
  Catalunya, Barcelona, Spain.
• Presenter Chen, Hung-Bin

ICASSP 2006
2
Outline

• Introduction
• On-line versions of the mean and variance
  normalization(MVN)
• Experiments
• Conclusions

3
Introduction

• On real-time system, However, the off-line
  estimation (MVN) involves a long delay that is
  likely unacceptable
• In this paper, we report an empirical
  investigation about several on-line versions of
  the mean and variance normalization(MVN)
  technique and the factors affecting their
  performance
• Segment-based updating of mean variance
• Recursive updating of mean variance

4
INVOLVED IN REAL-TIME MVN

• Mean and variance normalization

5
on-line versions issues

• Segment-based updating of mean variance
• there is a delay of half the length of the window
• Recursive updating of mean variance
• initialized using the first D frames of the
  current utterance and then they are recursively
  updated as new frames arrive

6
EXPERIMENTAL SETUP AND BASELINE RESULTS

• A subset of the office environment recordings
  from the Spanish version of the Speecon database
• carry out the speech recognition experiments with
  digit strings
• The database includes recordings with 4
  microphones
• a head-mounted close-talk (CT)
• a Lavalier mic
• a directional mic situated at 1 meter from the
  speaker
• an omni-directional microphone placed at 2-3
  meters from the speaker
• 125 speakers were chosen for training and 75 for
  testing
• both balanced in terms of sex and dialect

7
baseline system results

• the resulting parameters were used as features,
  along with their first- and second-order time
  derivatives

8
Segment-based updating of mean variance results

• a sliding fixed-length window centered in the
  current frame
• there is a delay of half the length of the window

9
Recursive updating of mean variance results

• Table 3 shows the results using different
  look-ahead values in the recursive MVN
• The entire look-ahead interval is used to
  calculate the initial estimates of mean and
  varianc
• in the case of 0 sec, the first 100 ms are used

10
Recursive updating of mean variance results

• The initial estimates were computed as in UTT-MVN
• This reinforces the good initial estimates of
  mean variance
• In our case ß was experimentally set to 0.992

11
INITIAL MEAN VARIANCE ESTIMATED FROMPAST DATA
ONLY

• this category do not use the current utterance to
  compute the initial estimates of the mean
  variance of the features
• the different data sources
• Current session
• in the current session with fixed microphones,
  environment and speaker
• utterances not included in the test set
• Set of sessions
• using utterances from a set of sessions of the
  Speecon database instead of utterances from the
  same session

12
initial with past data results
13
Conclusions

• In that case, a recursive MVN performs better
  than a segment-based MVN
• we observed that
• the usefulness of mean and variance updating
  increases when the initial estimates are not
  representative enough for a given utterance

14
GENERALIZED PERCEPTUAL FEATURES FOR VOCALIZATION
ANALYSIS ACROSSMULTIPLE SPECIES

• Patrick J. Clemins, Marek B. Trawicki, Kuntoro
  Adi, Jidong Tao, and Michael T. Johnson
• Marquette University
• Department of Electrical and Computer Engineering
• Speech and Signal Processing Lab
• Presenter Chen, Hung-Bin

ICASSP 2006
15
Outline

• Introduction
• The greenwood function cepstral coefficient
  (gfcc) model
• Experiments
• Conclusions

16
Introduction

• This paper introduces the Greenwood Function
  Cepstral Coefficient (GFCC) and Generalized
  Perceptual Linear Prediction (GPLP) feature
  extraction models for the analysis of animal
  vocalizations across arbitrary species
• Mel-Frequency Cepstral Coefficients (MFCCs) and
  Perceptual Linear Predication (PLP) coefficients
  are well-established feature representations for
  human speech analysis and recognition tasks
• W e want to generalize the frequency warping
  component across arbitrary species, we build on
  the work of Greenwood

17
Greenwood function cepstral coefficient (gfcc)
model

• Greenwood found that many mammals perceived
  frequency on a logarithmic scale along the
  cochlea
• He modeled this relationship with an equation of
  the form where a, A, and k are species-specific
  constants and x is the cochlea position
• This equation can be used to define a frequency
  warping through the following equations for real
  frequency, f, and perceived frequency, fp

18
Greenwood function cepstral coefficient (gfcc)
model

• Assuming this value for k, the other two
  constants can be solved for given the approximate
  hearing range (fmin fmax) of the species under
  study
• By setting Fp(fmin) 0 and Fp (fmax) 1, the
  following equations for a and A are derived
• Thus, using the species specific values for fmin
  and fmax and an assumed value of k0.88, a
  frequency warping function can be constructed

19
Greenwood function cepstral coefficient (gfcc)
model

• Figure 1 shows GFCC filter positioning for
  African Elephant and Passeriformes (songbird)
  species compared to the Mel-Frequency scale

20
Greenwood function cepstral coefficient (gfcc)
model

• Figure 2 shows GPLP equal loudness curves for
  African Elephant and Passeriformes (songbird)
  species compared to the human equal loudness
  curve

21
experiments

• Speaker independent song-type classification
  experiments were performed across the 5 most
  common song types using 50 exemplars of each
  song-type, each containing multiple song-type
  variants.

22
experiments

• Song-type dependent speaker identification
  experiments were performed using 25 exemplars of
  the most frequent song-type for each of the 6
  vocalizing buntings.

23
experiments

• Call-type dependent speaker identification
  experiments were performed using the entire
  Loxodonta Africana dataset
• There were a total of six elephants with 20, 30,
  14, 17, 34, and 28 rumble exemplars per elephant

24
Conclusions

• New feature extraction models have been
  introduced for application to analysis and
  classification tasks of animal vocalizations.