Sonority as a Basis for Rhythmic Class Discrimination Antonio Galves, USP. Jesus Garcia, USP. Denise Duarte, USP and UFGo. Charlotte Galves, UNICAMP.

Slides:



Advertisements
Similar presentations
An Adaptive, Dynamical Model of Linguistic Rhythm Sean McLennan GLM
Advertisements

An Adaptive, Dynamical Model of Linguistic Rhythm Sean McLennan Proposal Defense
How does first language influence second language rhythm? Laurence White and Sven Mattys Experimental Psychology Bristol University.
Spike Train Statistics Sabri IPM. Review of spike train  Extracting information from spike trains  Noisy environment:  in vitro  in vivo  measurement.
EP and BP Rhythm: Acoustic and Perceptual Evidence Sónia Frota Universidade de Lisboa Marina Vigário, Fernando Martins.
: Recognition Speech Segmentation Speech activity detection Vowel detection Duration parameters extraction Intonation parameters extraction German Italian.
A comparison of rhythms in Jamaican Creole speech and reggae music Project’s long term goals We chose to compare the rhythmic patterns of Jamaican Creole.
The statistical analysis of acoustic correlates of speech rhythm.
Languages’ rhythm and language acquisition Franck Ramus Laboratoire de Sciences Cognitives et Psycholinguistique, Paris Jacques Mehler, Marina Nespor,
Automatic identification of vocalic intervals in speech signal Jesus Garcia Antonio Galves Flaviane Fernandes Janaisa Viscardi Ulrike Gut Phonological.
Multiple testing Justin Chumbley Laboratory for Social and Neural Systems Research Institute for Empirical Research in Economics University of Zurich With.
1. Estimation ESTIMATION.
Statistics for Business and Economics
Causal Comparative Research: Purpose
Lecture 24: Thurs. Dec. 4 Extra sum of squares F-tests (10.3) R-squared statistic (10.4.1) Residual plots (11.2) Influential observations (11.3,
1 Simple Linear Regression Chapter Introduction In this chapter we examine the relationship among interval variables via a mathematical equation.
Business Statistics - QBM117 Interval estimation for the slope and y-intercept Hypothesis tests for regression.
Aaker, Kumar, Day Seventh Edition Instructor’s Presentation Slides
Scalable Text Mining with Sparse Generative Models
Today Concepts underlying inferential statistics
1 BA 555 Practical Business Analysis Review of Statistics Confidence Interval Estimation Hypothesis Testing Linear Regression Analysis Introduction Case.
Correlation & Regression
Review of Probability.
EE513 Audio Signals and Systems Statistical Pattern Classification Kevin D. Donohue Electrical and Computer Engineering University of Kentucky.
Statistical Methods For Engineers ChE 477 (UO Lab) Larry Baxter & Stan Harding Brigham Young University.
1. An Overview of the Data Analysis and Probability Standard for School Mathematics? 2.
Hypothesis Testing.
Sampling. Concerns 1)Representativeness of the Sample: Does the sample accurately portray the population from which it is drawn 2)Time and Change: Was.
As a conclusion, our system can perform good performance on a read speech corpus, but we will have to develop more accurate tools in order to model the.
BPS - 3rd Ed. Chapter 211 Inference for Regression.
Statistics for Business and Economics Chapter 10 Simple Linear Regression.
Infant Speech Perception & Language Processing. Languages of the World Similar and Different on many features Similarities –Arbitrary mapping of sound.
1 Institute of Engineering Mechanics Leopold-Franzens University Innsbruck, Austria, EU H.J. Pradlwarter and G.I. Schuëller Confidence.
1 Speech Perception 3/30/00. 2 Speech Perception How do we perceive speech? –Multifaceted process –Not fully understood –Models & theories attempt to.
© 2001 Prentice-Hall, Inc. Statistics for Business and Economics Simple Linear Regression Chapter 10.
MUMT611: Music Information Acquisition, Preservation, and Retrieval Presentation on Timbre Similarity Alexandre Savard March 2006.
Improving Speech Modelling Viktoria Maier Supervised by Prof. Hynek Hermansky.
Chapter 4 Linear Regression 1. Introduction Managerial decisions are often based on the relationship between two or more variables. For example, after.
6 - 1 © 1998 Prentice-Hall, Inc. Chapter 6 Sampling Distributions.
1 Statistical NLP: Lecture 7 Collocations. 2 Introduction 4 Collocations are characterized by limited compositionality. 4 Large overlap between the concepts.
A Model for Learning the Semantics of Pictures V. Lavrenko, R. Manmatha, J. Jeon Center for Intelligent Information Retrieval Computer Science Department,
Experiments on Noise CharacterizationRoma, March 10,1999Andrea Viceré Experiments on Noise Analysis l Need of noise characterization for  Monitoring the.
Conditional Random Fields for ASR Jeremy Morris July 25, 2006.
MGT-491 QUANTITATIVE ANALYSIS AND RESEARCH FOR MANAGEMENT OSMAN BIN SAIF Session 22.
Chapter 20 Classification and Estimation Classification – Feature selection Good feature have four characteristics: –Discrimination. Features.
Spatial Smoothing and Multiple Comparisons Correction for Dummies Alexa Morcom, Matthew Brett Acknowledgements.
Chapter 6 Conducting & Reading Research Baumgartner et al Chapter 6 Selection of Research Participants: Sampling Procedures.
Review of Statistical Terms Population Sample Parameter Statistic.
Education 793 Class Notes Inference and Hypothesis Testing Using the Normal Distribution 8 October 2003.
Testing a Single Mean Module 16. Tests of Significance Confidence intervals are used to estimate a population parameter. Tests of Significance or Hypothesis.
Essential Statistics Chapter 171 Two-Sample Problems.
6 - 1 © 2000 Prentice-Hall, Inc. Statistics for Business and Economics Sampling Distributions Chapter 6.
Jump to first page Inferring Sample Findings to the Population and Testing for Differences.
Statistical Inference for the Mean Objectives: (Chapter 8&9, DeCoursey) -To understand the terms variance and standard error of a sample mean, Null Hypothesis,
Educational Research Inferential Statistics Chapter th Chapter 12- 8th Gay and Airasian.
BPS - 5th Ed. Chapter 231 Inference for Regression.
Chapter 6 Sampling and Sampling Distributions
Class Six Turn In: Chapter 15: 30, 32, 38, 44, 48, 50 Chapter 17: 28, 38, 44 For Class Seven: Chapter 18: 32, 34, 36 Chapter 19: 26, 34, 44 Quiz 3 Read.
17th International Conference on Infant Studies Baltimore, Maryland, March 2010 Language Discrimination by Infants: Discriminating Within the Native.
Advanced Higher Statistics
Inference for Proportions
Chapter 11: Simple Linear Regression
Behavioral Statistics
Sampling Distribution
Sampling Distribution
EE513 Audio Signals and Systems
Research in Psychology
Speech recognition, machine learning
Speech recognition, machine learning
Presentation transcript:

Sonority as a Basis for Rhythmic Class Discrimination Antonio Galves, USP. Jesus Garcia, USP. Denise Duarte, USP and UFGo. Charlotte Galves, UNICAMP.

The starting point : Ramus, Nespor & Mehler (1999)

What we do Our goal: a new approach to the problem of finding acoustic correlates of the rhythmic classes. Main ingredient: a rough measure of sonority defined directly from the spectrogram of the signal. Major advantage: can be implemented in an entirely automatic way, with no need of previous hand-labelling of the acoustic signal.

Our main result Applied to the same linguistic samples considered in RNM, our approach produces the same clusters corresponding to the three conjectured rhythmic classes.

RNM revisited Striking features Linear correlation between ΔC and %V (-0.93). Clustering into three groups.

A parametric probabilistic model for RNM Duarte et al. (2001) propose a parametric family of probability distributions that closely fit the data in RNM. This has two advantages: It provides a deeper insight of the phenomena. It makes it possible to perform statistical inference, i-e to extend results from the sample (data set) to the population (the set of all potential sentences).

The probabilistic model The duration of the successive consonantal intervals are independent and identically distributed random variables. The duration of each consonantal interval is distributed acording to a Gamma distribution. Different languages have Gamma distributions with different standard deviations. The standard deviation is constant for all languages belonging to the same rhythmic class. The standard deviations of different classes are different.

Statistical evidence for the clustering The model enables testing the hypothesis that the eight languages are clustered in three groups. The hypothesis that the standard deviations of the Gamma distributions are constant within classes and differ among classes are compatible with the data presented in RNM.

Estimated standard deviations of the Gamma distribution for the consonantal intervals

Problems for RNM (1) RNM is based on a hand-labeling segmentation which is time-consuming and depends on decisions which are difficult to reproduce in an homogeneous way. This is a problem for linguists.

Problems for RNM (2) Newborn babies discriminate rhythmic groups from signal filtered at 400 Hz (Mehler et al. 1996). At this frequency, it is impossible to fully discriminate consonants and vowels. ΔC depends on a complex computation. This is a problem for babies!

Sonority as a basis for rhythmic class discrimination Mehler et al. (1996)’s results strongly suggest that the discrimination of rhythmic classes by babies relies not on a fine-grained distinction between vowels and consonants, but on a coarse-grained perception of sonority in opposition to obstruency. A natural conjecture is that the identification of rhythmic classes must be possible using a rough measure of sonority.

A rough measure of sonority Goal: to define a function that maps local windows of the signal on the interval [0,1]. This function should assign values close to 1 for spans displaying regular patterns, characteristic of the sonorant regions of the signal, values close to 0 for regions characterized by high obstruency.

Technical specifications The function s(t) is based on the spectrogram of the signal. Values of the spectrogram are estimated with a 25ms Gaussian window. The step unit of the function is 2ms. Computations are made with Praat (

Definition of the function s(t) p t (f) = re-normalized power spectrum for frequency f around time t. This re-normalization makes p t a probability measure. A regular pattern characteristic of sonorant spans will produce a sequence of probability measures which are close in the sense of relative entropy. This suggests defining the function sonority as

Values of 1- s(t) on a Japanese example

Estimators

Explaining the estimators is the sample mean of the function s(t). δS measures how important are the high obstruency regions in the sample. This is due to the fact that typically the values of p(t), and consequently s(t), present large variations when t belongs to intervals with high obstruency.

Distribution of the eight languages on the (,  S ) plane

Extra statistical features LanguagesP(s<0.3)Q3-Q1 Japanese Catalan Italian Spanish French Polish English Dutch The distance between the first and third quartile increases from Japanese to Dutch. In other terms, the dispersion of sonority increases from mora-timed to stress-timed languages. The empirical probability of having sonority smaller than 0.3 also increases from Japanese to Dutch. This reinforces the idea present in Duarte et al. (2001) that the relevant information to discriminate among rhythmic classes is contained in the less sonorant part of the signal.

Distribution of the eight considered languages on the (,%V) plane

Distribution of the eight languages on the (  S,  C) plane

Conclusions The main purpose of this presentation was to show that the relevant evidence about rhythmic classes can be automatically retrieved from the acoustic signal, through a rough measure of sonority. In addition, our statistics are based on a coarse-grained treatment of the speech signal which is likely to be closer to the linguistic reality of the early acquisition.

This work is part of the Project RHYTHMIC PATTERNS, PARAMETER SETTING AND LANGUAGE CHANGE, funded by Fapesp (grant no 98/ ).

Values of 1- s(t) on a Dutch example

Values of 1- s(t) on a French example