study of word-level accent classification and gender factors xing wang, peihong guo, tian lan, guoyu...

15
Study of Word-Level Accent Classification and Gender Factors Xing Wang, Peihong Guo, Tian Lan, Guoyu Fu CSCE 666 Term Project Presentation Dec 11th, 2013

Upload: barry-gregory

Post on 27-Dec-2015

214 views

Category:

Documents


0 download

TRANSCRIPT

Study of Word-Level Accent Classification and Gender Factors

Xing Wang, Peihong Guo, Tian Lan, Guoyu Fu

CSCE 666Term Project Presentation

Dec 11th, 2013

Background• Motivation: Accent Recognition(AR) helps improve

Speech Recognition system and Speaker Identification system

• Our worko Word-level Classifiers are built using different types of features, words

and learning methodso Speaker variation affect the AR system, gender factor is considered in

this work.

Outline

• Feature Processingo Word Alignmento MFCC, Formants

• Classifiero GMM, HMM

• Experimentso Comparison of features, classifiers and wordso Gender effect

Data preparation• Audios and corresponding phonetic

transcriptions• Biographical data for speakers

o Web crawler (Python)

• All metadata stored in Databaseo Faster to locate and extract audio information

Feature

Processing

Alignment• The Penn Phonetics Lab Forced Aligner is used

to segment words apart.• We are doing the word-based accent

recognition system.

Feature

Processing

Feature Extraction• Frame

o Window size, 25mso Window shift, 10ms

• MFCCs

• Formants

Feature

Processing

Classifiers

• GMMo Number of components are determined via cross validationo EM Algorithm to train

• HMMo Observation: MFCC or Formantso Hidden states: Single Gaussian componento EM algorithm for training

Comparison of Features• Experiments Process

1. 5-fold cross validation2. Training and test3. Repeat 5 times on random samples

• HMM is fixed as the classifier

• Features for comparison1. MFCCs (c0~c12)2. F0F1F23. F1F2

Experiments

Comparison of Classifiers• MFCC is fixed as the approach to extract

features

• Classifiers for comparison1. Non-temporal: GMM2. Temporal: HMM

Experiments

Comparison of Words• Factors of classification accuracies of

different wordso Certain vowels and consonantso C1-C2 trajectory of word ‘OF’

Experiments

Gender EffectGender classification

Experiments

• Thanks!•Q&A