Scaling Studies of Perceived Source Width Juha Merimaa Institut für Kommunikationsakustik Ruhr-Universität Bochum

Binaural and Spatial Hearing Group

Scaling Studies of Perceived Source Width

Juha Merimaa

Institut für KommunikationsakustikRuhr-Universität Bochum


Outline

• Introduction• Background on listening tests• Description of the conducted pilot test• Analysis methods & preliminary results• Discussion & summary


Introduction

• A room or a hall broadens the perceived width of auditory objects

• Traditionally auditory source width (ASW) has been investigated as a descriptor for concert halls

• How does the broadening depend on source signals?


In other words...

• In a scene based paradigm– source broadening is due to the part of room

effect that is grouped with source signals– the rest of room effect is resolved into a

separate percept

• What are the spatial features related to auditory “deconvolution” of reverberation


Listening test basics

• Quantifying auditory perception• Levels of measurement

Interval

OrdinalNominal

Ratio

Short Long 1 2 3

1 2 30 41 2 30 4


Possible test methods for assessing ASW

• Direct scaling– Rating– Rank ordering– Assigning stimuli in successive categories

• Constant reference– All stimuli are judged relative to a single

reference stimulus


Possible test methods (contd.)

• Method of adjustment– Listeners adjust a variable reference to

correspond to each stimulus

• Adaptive procedures– Reference is adaptively adjusted based on

listeners judgements

• Pairwise comparisons– Each stimulus is judged relative to all others


Why pairwise comparisons?

• Source broadening is expected to be a sum of several interaural signal features

• All except pairwise comparison methods force the results onto a linear scale– Weighting of dimensions implicit in the data• Can be accessed with factor analysis

– Weights may vary between individuals, which will result in noisy unidimensional data


Pilot listening test• Gathering both preference and distance

data between pairs


Stimuli

– Speech (sp)

– Cello, f0 = 196 Hz

(ce)– Snare drum (sn)

– Two harmonic complexes,f0 = 196 Hz, -12 dB/oct• No modulation (h1)• Freq. mod. 1%, 6 Hz (h2)

– Pink noise 100 Hz –10 kHz (ns)

• Anechoic samples convolved with binaural room responses


Binaural room responses

• Diffuse field and system compensated responses– Medium size diffuse concert hall (p)• RT = 2.2 s, 1-IACC

E3 = 0.78

– Large multipurpose hall (a)• RT = 2.4 s, 1-IACC

E3 = 0.02

– Small listening room (s)• RT = 0.5 s, 1-IACC

E3 = 0.32


Altogether 18 stimuli resulting in 153 permutations


Analysis of preference data

• A single run comparing all the the pairs results in a preference matrix that can be used to rank order the stimuli

• In an ideal case each run will yield the same perfectly ordered set of data

A B C DAB 1C 0 0D 1 1 1


Real world comparative judgements

• Each stimulus has a dispersion on a psychological scale

• Each judgment of distance and order depend on current points of perception


Checking for consistency

• Circular triads• Mean for random answers

with 18 stimuli: 204• Average in collected data approx. 40• All data matrices consistent with

significance p < 0.01

A B

C


Unidimensional scaling

• Simplest scaling method: count the number of times a single stimulus is prefered over all others


Wincount statistics, CR = 60

0 20 40 60 80 100 120 140

s_h1s_h2a_h1a_sns_sna_h2s_cea_cea_nsa_sps_sps_nsp_snp_cep_h1p_spp_h2p_ns


Comparison with stimulus IACC

0 20 40 60 80 100 120 140

s_h1s_h2a_h1a_sns_sna_h2s_cea_cea_nsa_sps_sps_nsp_snp_cep_h1p_spp_h2p_ns

Wincount

0 0.15 0.3 0.45 0.6 0.75 0.9 1.051-IACC

3


More sophisticated scaling

• Thurstone's law provides a method for mapping pair comparison data on an interval scale– Assumes normally distributed

unidimensional data– Includes tests for checking the fit

• Results– Significance of deviation from data p < 0.01


Multidimensional scaling

• Uses distances between stimuli to construct a spatial representation ofdata in n dimensions

• Metric (interval) and nonmetric (ordinal) procedures

• Few assumptions on data• Works well with a relatively small number

of test subjects


-2-1

01

23 -2

0

2-1.5

-1

-0.5

0

0.5

1

1.5

p_ns

p_sp

p_h1

p_h2

p_ce p_sn

s_sp

s_ns

a_ns

a_sp

a_sn a_h2

a_ce

s_sn s_ce

s_h2

s_h1

a_h1

3-D scaling of all stimuli


Comparison of the large halls

-2 -1 0 1-2

-1.5

-1

-0.5

0

0.5

1

1.5

2

a_ce

a_h1 a_h2

a_ns

a_sn

a_sp

p_ce p_h1

p_h2

p_ns

p_sn

p_sp


Multipurpose hall vs. Listening room

-2 -1 0 1-2.5

-2

-1.5

-1

-0.5

0

0.5

1

1.5

a_ce

a_h1

a_h2

a_ns

a_sn

a_sp

s_ce

s_h1

s_h2

s_ns

s_sn

s_sp


Concert hall vs. Listening room

-2 -1 0 1 2 -2

0

2

-2

-1.5

-1

-0.5

0

0.5

1

1.5

2

p_h1 p_h2

p_ns

p_sp

p_ce

p_sn

s_ns

s_sp

s_sn

s_ce s_h2

s_h1


Discussion & conclusions

• The perception of auditory source width is clearly multidimensional– Results between the most similar spaces

suggest separate source and room dimensions with some interaction

– Euclidian metric of MDS might not reflect human perception between extreme cases

• The pilot data is insufficient to draw more firm conclusions


Future work

• A larger listening test with a reduced set of stimuli

• Interpreting the dimensions in terms of binaural cues– Breaking the experiment into several

unidimensional studies– Use gained results in choosing stimuli

• Similar investigations into envelopment

Documents

Scaling Studies of Perceived Source Width Juha Merimaa Institut für Kommunikationsakustik Ruhr-Universität Bochum