Pronunciation Modeling of Mandarin Casual Speech

8/24/00

Click here to start

Table of Contents

Pronunciation Modeling of Mandarin Casual Speech

Project Objectives 

CASS Corpus Review 

CASS Corpus Review 

Chinese Segmental Labeling Convention SAMPA-C 

Generalized Initial/Final (GIF)

Variability in the CASS Corpus 

PPT Slide

Overview Map (GIF/IF Modeling on CASS) 

Overview Map (Decision Tree Approach)

Flexible Alignment Tool

Flexible Alignment Tools

PPT Slide

PPT Slide

PPT Slide

PPT Slide

Extensibility of the Aligner

GIF/IF Pronunciation Modeling 

GIF/IF Sparse Data

Probabilistic GIF

State Level Gaussian Sharing 

State Level Gaussian Sharing

State Level Gaussian Sharing

Deleted Interpolation of IF/GIF Models

Deleted Interpolation of IF/GIF Models

PPT Slide

Discussion of Experimental Results

GIF/IF Pronunciation Modeling 

Probabilistic Pronunciation Models

Acoustic Modeling of IF/GIF with Sparse Data: P(a|k, s)

Using Context Equivalence Classes for P(s|k)

Experimental Results

Comparison Among Equal-Probability, DOP, and CDW 

Discussion of Experimental Results

On going work: Pronunciation Modeling with Sparse GIF Data

Predictive Features for Pronunciation Modeling 

Sound Changes Annotated in CASS

Predicting Voicing Directly

High Correlation with Linguistic Transcription of Voicing Feature 

Using Voicing

Training Pronunciation Models on CASS

Training Pronunciation Models on CASS

Incorporation of Acoustic Features into Pronunciation Models 

Deriving Regular Phone Change Patterns From The CASS Corpus 

Example `Artificial' Phone Change 

Apply CASS Pronunciation Models to Broadcast News Data

Mandarin Broadcast News Evaluation

Ongoing Work: Inching Towards GIFs 

Alternatives to Explicit GIF Models 

Example (real) Triphone Clustering Trees Using Acoustic Side Information

End of Presentation

Important Points 

Important Points

Important Points 

Important Points