Testing the perception of speech in noise in children: A test with no verbal response Intelligibility and Quality of Speech in Noise Workshop, UCL 9 th.

Testing the perception of speech in noise in children: A test with no verbal response

Intelligibility and Quality of Speech in Noise Workshop, UCL

9th January 2009

Sam Evans & Stuart RosenDept of Speech, Hearing & Phonetic Sciences, UCL

Perception of Speech in Noise in Children

• Noisy classrooms – academic, literacy and language development.

• Children with Specific Language Impairment (SLI) and dyslexia have difficulty listening in noise (Ziegler, 2005; Bradlow et al 2003; Ziegler, in press).

Speech in noise

Language & literacy

Motivation for the Development of a New Test

• Numerous speech in noise tests.• Focus at different levels of morphological and

syntactic complexity. • Few tests developed specifically for children. • Adult tests adapted for children without

considering needs.

Motivation for the Development of a New Test

• We are interested in SLI and dyslexia – perception of single words.

• Nonsense syllables – poor ecological validity.

Sentences – potential syntactic confound.

• No single word tests that meet our needs.

Specific Challenges for Designing a Test for Children with Language Difficulties

• Open response problematic – speech production errors.

• Difficulties often subtle – require minimal speech contrasts.

• Poor vocabulary - lexical confound.• Poor attention – quick to administer.

How we Addressed these Issues: Our Test

• Non verbal response - choose from a “triplet” of items: Target + two non word foils: e.g. “bite” “pite” “dite”.

• Foils differ from target by a single feature of voice, place or manner.

• Targets chosen from lists of the earliest acquired words.• Chance level (33%) – requires less items than 2AFC.

Adaptive procedure – negates floor and ceiling effects.

A Demonstration….

Construction of the Test: Choosing Targets

Each triplet met the following criteria:• CVCs• Acquired by age 3;6, based on estimations from the

databases of: Morrison, Chappell & Ellis (1997); Bird, Franklin & Howard (2001); DeCara & Goswami (2002).

• Could be represented pictorially • Altered by a single feature to derive two non-word foils.

Difficult to balance the contrasts perfectly: 7 manner-voice, 7 manner-place & 14 place-voice.

Equalising Across Triplets

• Each target has a different relative level of intelligibility due to phonemic and psycholinguistic factors (Nilsson et al 1994; Papso & Blood, 1989).

• A relative dB correction calculated for each triplet.• Correction factor applied to nominal SNR in the

adaptive procedure.

Equalising Across Triplets

Correction factor calculated by:

• Testing each triplet at a range of SNRs – 110 data points.• Logistic Regression - 71% correct SRT for each triplet.• Mean SRT for all triplets subtracted from the SRT of each

individual triplet.

Example Psychometric Functions for triplets containing “bite” and “boot”

• “bite” “dite” “pite”

Target Place Voice

SRT = -11.3 (71% correct)

• “boot” “woot” “poot”

Target Manner Voice

SRT=-0.7 (71% correct)

Adaptive Procedure

• 2-down/1-up adaptive procedure tracking 71% correct (Levitt, 1971)

• Final step size 3dB. • Familiarisation & test

stages.• SRT established by the

mean of the reversals in the test phase.

Imp

rovi

ng P

erfo

rman

ce

Speech Shaped Noise Masker

• Speech shaped noise was generated by estimating the long-term average speech spectrum from recent measurements for combined male and female voices (Table II of Byrne et al., 1994).

Result 1: Effects of Age

• 53 subjects without known hearing or learning impairments, aged from 5 to 25 yrs - speech shaped noise masker.

• An inverse curve accounted for the most variance using the least number of parameters.

• In keeping with previous findings showing poorer perception of speech in noise in children (Talarico et al 2007; Fallon, Trehub & Schneider, 2000; Papso & Blood, 1989).

SR

T (

dB

)

Imp

rovi

ng P

erfo

rman

ce

Result 2: Effects of Vocabulary Development

• 16 monolingual British English speaking children (mean age=11.12, SD=2.78, range=6.92-14.50).

• 26 children from a multilingual background (mean =10.32, SD=2.26, range=6.75-13.50).

• All children were without known hearing or learning impairments.

• Tested in free-field with a speech shaped noise masker.• Vocabulary was assessed with the Renfrew Word Finding

Vocabulary Test (raw scores are reported).

Result 2: Effects of Vocabulary Development

Vocabulary:• Significant effect of age and group. No age*group interaction.

SRT:• Significant effect of age, no effect of group or age*group interaction.

With thanks to Aneeka Degun

Imp

rovi

ng p

erfo

rman

ce

Imp

rovi

ng p

erfo

rman

ce

6.00 8.00 10.00 12.00 14.00

Age (Years)

-12.00

-10.00

-8.00

-6.00

-4.00

-2.00

SR

T (

dB

)

GroupMonolingual British English Speakers

Mutlilingual

SRT as a Function of Age

R Sq Linear = 0.335

R Sq Linear = 0.107

6.00 8.00 10.00 12.00 14.00

Age (Years)

25

30

35

40

45

50

Vo

ca

bu

lary

Ra

w S

co

res

GroupMonolingual British English Speakers

Mutlilingual

Vocabulary as a Function of Age

R Sq Linear = 0.529

R Sq Linear = 0.324

A Test Robust to Differences in Vocabulary Development

• Groups differed in vocabulary development but not in SRT, suggesting that we have developed a test which is not sensitive to differences in vocabulary development, at least within the ranges assessed.

Result 3: Effects of Practice

• 14 children (mean age=8.4,SD=1.82,range=5.4-11.1). 2 subjects excluded – poor performers.

• All subjects without known hearing or learning impairments.

• 4 different maskers. • Two separate sessions (up to 1 month between)

with each masker (randomised in order).

Maskers…

• Speech shaped Noise• One talker – speaker taken from EUROM database, edited

to exclude pauses more than 0.1s in duration.• Two talkers – the above talker plus an additional talker

again taken from the EUROM database.• Modulated noise – speech shaped noise multiplied by the

envelope of the one talker stimuli.

Practice Effects

• Control for masker: Each SRT observation was subtracted by the mean SRT of the masker type that it belonged to.

• Significant effect of age.• No significant effect of session,

order or interactions.• No evidence of practice effects

within or between sessions.1 2

Session

-20.00

-10.00

0.00

10.00

20.00

SR

T S

tan

dar

dis

ed b

y M

aske

r

10

85

23

26

91

98

108

Order1

2

3

4

With thanks to Henri Roe

Looking to the Future

• Norming with larger samples of children – in process.

• Analysis of Errors – Ziegler• Large SLI study at Guy’s Hospital – Gillian Baird

and colleagues.

Thank yous

With thanks to:

for the award of a vacation scholarship.

And all the students who have worked on/are currently working on the test:

Henri Roe, Aneeka Degun & Rebecca Lancaster