A Simplified System for Analyzing

Stop Consonant AcousticsBy Zach Polen

Advisor: Helen Hanson

BackgroundBackgroundCurrent study among Hanson (Union College),

Shattuck-Huffnagel(MIT) and Demuth(Brown) is being conducted to look at medial stops in kids.

Medial stops in this case are the consonants g,t,p,k,d and b

Why? Trying to develop models of speech development in children

What purpose does What purpose does this serve?this serve?

Provides insight into how language is stored and processed in the brain

Investigates how speech utterances are planned and produced

Possible applications to diagnose and treat speech disorders

Speech recognition and speech synthesis

Labeling Speech Events Labeling Speech Events In Children’s Speech In Children’s Speech

UtterancesUtterances

Compute Durations From Compute Durations From Labeled DataLabeled Data

Voice Bar

Vowel Duration

Onset noise

Vowel Final Noise

AnalysisAnalysis Compare Across

Subjects (10 children/mothers) Voicing ( voiced/ voiceless Place of Articulation (location on the tongue)

Current MethodCurrent MethodCurrent approach uses several programs in

order to get required data.

Acoustic Recordin

g

TextGrid Formed in

Praat

Script extracts labels/ti

mes

Excel computes durations/averages

Excel Data to .TXT

MATLAB statistical analysis

Output to Excel/Word

Current Process Current Process ExplainedExplained

1. The first step is taking the acoustic data and analyze it using TextGrids in Praat. This needs to be repeated for as many files as you have

2. The Computation/Measurement script has to be ran for as many different measurements as you are looking for. In order to do so you have to know which acoustic labels need to be subtracted from each other in the TextGrids just formed above. You also have to insert these values manually.(Created by research student at MIT)

3. The Averages and Standard Deviations then have to be found in Excel for each type of measurement you are looking for and which acoustic tokens you want to account for. (Done by Summer Research Student at Union)

4. The data needs to be put into a form MATLAB can read in, so in this case you take an Excel file and convert it to a text file. (Done during Summer Research by myself )

5. The data is read into MATLAB and put into a form in MATLAB so that it can run the statistical analysis. (Done during Summer Research by myself )

6. The data is not stored/saved in MATLAB so we have to take the data and tables and output them into either Word or Excel.

Problems with Current Problems with Current MethodMethod

Piece-meal method; although it works, leaves a lot of room for human error.

Examples: Uniform labeling system not enforced Requires numerous conversions vs. data

formats Excel analysis not automated Confusing/Complex

Researcher’s RequestsResearcher’s Requests

-Simplify system (time, process, # programs)

-Reduce opportunity for human error

-Enforce label uniformity

Design ConsiderationsDesign Considerations1. Semi-automate labeling of speech events

using Praat software

2. Import the Praat data into a MATLAB database

3. Perform the statistical analysis of data in MATLAB

4. Have the system flexible enough so that it can be used for other studies

Projected Block Projected Block DiagramDiagram

LabelingAnalysis/

Computations

Acousticrecordings

Labels/time

Tables of Statistics/Durations

Design AlternativesDesign Alternatives Current piece-meal system

Single program that: Label waveforms, read in data, analyze data,

output statistics

Hybrid system Praat for labeling acoustic waveforms MATLAB for extracting/analyzing/outputting

data

Final Design: Final Design: Implement Hybrid Implement Hybrid

ProcessProcess Completely redoing entire system would

require more time.

Why we chose to use MATLAB: Experience Most familiar programming language for both

student/advisor Users have access

DesignDesign First we had to decide how we wanted to

save the data in MATLAB.

We decided that we wanted to create a database structure.

This way the data could be conveniently stored and accessed

Breaking Down the Breaking Down the TitlesTitles

From the textgrid filenames we were able to get a lot of classification information for our database: ID Playsession Word Token Place Sound/type Voicing

Getting TextGrid Getting TextGrid InformationInformation

Ran a succession of while loops

Outer most loop contained a for loop which went through all the records

First while loop found the number of tiers

The next while loop then processes each tier (which has another 4 while loops embedded) First finds name of tier Second finds number of points in tier Third finds label of the speech event Fourth finds the time associated with that label

BreakdownBreakdown# Tiers While Loop

Number Tiers

Tiers Processed While

Loop

Name Tier

Size While Loop

Num Points

Point Process

Loop

Name Point

Time at Point

TextGridCycle

Through Records

Loop

Directory of Records

Troubleshooting Troubleshooting problemsproblems

Problem with the manual labeling is even if using the exact same format you can still make mistakes. In order to account for this we had checks in MATLAB that would make all the words lower case (if label was supposed to be v1_begin and you had V1_BEgIn MATLAB would auto correct), removed symbols like “?”(if you had made the mistake of say V1-?begin MATLAB would correct to v1_begin) and made sure all “-” were “_”.

How to save our data?How to save our data? We wanted to dynamically save our data

Allows for data on the spot This way we don’t just get last result of each

loop Get the data saved for each iteration of the for

loop Data still accessible and usable later Saving was not dependant on using our

method but could be done with other studies

The Answer?The Answer?An “eval” statement!!!

This was harder to conceptualize because you do not necessarily see what it is doing as clearly as a longer code which could save it.

The eval statement takes a bunch of strings/numbers and executes them the way you have it written out.

Sample:

eval([DurType{p} '_avg' '.' alist{m} '.' slist{i} '.' vlist{n} '.' plist{l} '=' 'mean(SAVP);'])

Another great thing Another great thing about eval!about eval!

We could have saved the data and created our database using a bunch of switch statements, or for loops, etc… But had we done it this way it would be specific to our study

The eval lets you have your list of whatever durations/events you want and from that list executes each specific one.

So you could have the label names a, b , c and I could have d, e, f and either way the eval will correlate and find the desired measurements

Duration ComputationDuration ComputationUsing a for loop and our saved database

structure we then used if statements to derive computations

What if speech events weren’t marked?

Then we found out the unique labels for subject, age, voicing and place (what we want to average across)

Duration Duration Computations Cont.Computations Cont.

We wanted to average across token in subject group:

Needed to find where subject and age intersected

Found where subject,age,voicing intersected: averaged durations across place

Found where subject, age, voicing, place intersected: averaged durations across token

Dynamically saved using eval statements

Duration Duration Computations Cont.Computations Cont.

We then wanted to average across subject in age group:

Found where age, voicing intersected: averaged across subject/place

Found where age, voicing, place intersected: averaged across subject

ComplicationsComplicationsWhat if there was no events for a given

duration?

find locations

remove locations

Why?

What else we’d like to What else we’d like to get done…get done…

Create a script which outputs durations in a neatly formatted table

Create simple GUI that allows for user to pick which durations they want formatted in table

Using our durations create a script to run standard deviations and statistical (ANOVA) analysis

Praat script which automatically puts in the labels wanted for textgrid (eliminates hand error of writing labels, and keeps all labeling uniform)

Advantage of Our Advantage of Our MethodMethod

We said we want to reduce time without reducing accuracy and hopefully even increase accuracy!

Our process reduces time which can be seen from the block diagram. There are two steps the user must take versus 6 steps in the old process.

Accuracy improves because there is less manual involvement and more automation. There are also checks in the code which eliminate previously made errors so the accuracy is even more improved.

Data does not go from program to program so less chance of a copying error. Formula’s for statistical analysis are also already input so the user does not have to worry about making a mistake on those in Excel.

All the user has to do is label their textgrids and the outputs they are looking for will be obtained.

Future WorkFuture Work Developing an automated labeling script

Simplify MATLAB code into a series of functions

AcknowledgementsAcknowledgements

Questions?Questions?