Acoustic-Articulatory Interpretations of Professional Imitations of “hallå” in three Swedish Dialects Frantz Clermont & Elisabeth Zetterholm Department.

Acoustic-Articulatory Interpretations of Professional Imitations of “hallå”

in three Swedish Dialects

Frantz Clermont & Elisabeth ZetterholmDepartment of Linguistics & Phonetics

Centre for Languages & LiteratureLund University, SWEDEN

2006 IAFPA Conference Gothenburg University, Sweden

25th July 2006

2

A Forensic Scenario: Hypothetical but Not Implausible

Managing Director of the Bank of Sweden

receives 5 threatening calls on the same day

A male voice perceived for all 5 calls

A different dialect perceived on each call

Every call followed a similar pattern: the caller(s) started by saying “hallå”

politely with situational variations

the caller(s) proceeded with threats

expressed in harsh language

The EVIDENCE FORENSICALLY-ORIENTED QUESTIONS

1a) What dialectal markers are detectable

in “hallå” spoken by a given imitator?

1b) Are the markers consistent

from situation to situation?

2a) Are there cues of an imitator’s

in dialectal imitations?

2b) Do such cues persist

from situation to situation?

Clermont, F. & Zetterholm, E. (2006), “F-pattern Analysis of Professional Imitations of “hallå”

in three Swedish dialects”, Proc. XIXth Swedish Phonetics Conf., 7-9 June 2006, Lund.

THIS SEQUEL STUDY:

ACOUSTIC-ARTICULATORY INSIGHTS

OBSERVATION from F-patterns:“LANDMARKS of RESISTANCE”

in NEIGHBOURHOOD of /l/ in “hallå”

3

(1) To detect dialectal markers in “hallå” 1 situation (telephone-answering)

1 professional imitator

3 of the 5 dialects recorded

(2) To look for cues of the imitator’s voice

in the 3 dialectal imitations

Data Particulars & Research Objectives

meeting along-lost friend

SITUATIONS

answering the telephone

signallingarrival at home

“hallå, det är Anders”

“hallå, nu är jag hemma”

“hallå, det var längesen”

NorrlandIMITATOR

DIALECTAL IMITATIONS

Stockholm Gothenburg Skåne Småland

STKSLF GTB SKN SLD NLD

øø ø ø ø ø

6 6

7 7

6

6 6

6 6

6 6 6 5

7 7

6 6 6 7 5 5

5 56 6

Listening Test Results on Impressionistic Scale

[v. poor = 1 2 3 4 5 6 7 = v. good]

ACOUSTIC-ARTICULATORY

APPROACH

4

Phonetic

Segmentation

Time

Normalisation

F-pattern

Estimation

POLE-GRAMS

‘a’ ‘l’ ‘å:’

‘a’ ‘l’ ‘å:’

Acoustic-Articulatory Methods: (1) Segmentation, F-pattern

?

The Problem!

5

Acoustic-Articulatory Methods: (2) Time Normalisation[Approach guided by relative durations of ‘a’, ‘l’, ‘å’]

/l/-segment - a suitable reference segment Rule of thumb:

L(/a/) ≈ 3 times L(/l/)

L(/å:/) ≈ 5 times L(/l/) 5 landmarks for /l/ 15 for /a/ & 25 for /å:/

a total of 45 landmarks Values @landmarks via spline interpolation

‘a’ ‘l’ ‘å:’

Inter-token spread (Hz) averaged over the 45 landmarks

F1 F2 F3 F4

mean (+SLF)Mean (−SLF)

24 24

4442

47 31

49 48

6

Acoustic-Articulatory Methods: (3a) SM67-Model

SM67-Model Formant frequencies Anti-symmetric component of = odd-numbered terms of a Fourier cosine series = even-numbered terms not easily obtained from

acoustics

Anti-Symmetry A limitation vis-à-vis symmetrically-shaped sounds An advantage vis-à-vis shape uniqueness

EXTANT PROBLEMS How unique is the SM67’s Area Function? How to set a realistic value for L?

10 A0nnn FFF

L

cnFn 4

120

SM67: [Schroeder (1967); Mermelstein (1967)]

{F1, F2, F3, …}

L

dx xAln

xAln12 na

na2

Sound Speed

Acoustic-Articulatory Methods: (3b) SM67-Model extended

SynthesiseCalculate12 na xAln

Calculate

Cosine Coefficients xATARGETF-Pattern

CANDIDATEF-Pattern

iterate untilerror ≤ 0.001

adjust 12 na

SM67SM67 LF75

Calculate ECCENTRICITYfrom uniformity

adjust L = L ± dL Lmin = L0−3cm Lmax = L0+3cm dL = 0.5cm

i

2i

i

i

1)(2iF

1)(2iF

4

cL0

PZ70

012 2

n

nn F

Fa

Parseval’s Theorem

212 )a( n

n

7

‘a’

‘l’

‘å’

Acoustic-Articulatory Representation of “hallå”:

Examples of Key Parameters

landmark 7

landmark 18

landmark 30

Eccentricity from Uniformity

over a range of candidate VT-Lengths

Cross-Sectional Area Function (@min. eccentricity)

optimised to match target F-patterns

8

Vocal-Tract Shapes (area functions) & Lengths

SELF

GBG STK SKN

DIALECTAL landmarks

Potential Landmarks

of RESISTANCE

SELF

9

Vocal-Tract Shapes & Eccentricity from Uniformity

SELF

GBG STK SKN

Quasi-Neutral

Landmarks

DIALECTAL landmarks

10

Vocal-Tract Shapes & /l/-related CoarticulationRMS distances between

Area-Functions of ‘l’ (onset-landmark) and ‘Area-Functions of a’ (15 landmarks)

RMS distances between

Area-Functions of ‘l’ (offset-landmark) and Area-Functions of ‘å’ (25 landmarks)

Coarticulation-Resistance

Landmarks

11

Concluding Summary

Motivated present endeavour by way of a forensic “hypothetical”

Outlined current project on dialectal imitations of “hallå”

Discussed time-normalisation problem & demonstrated approach

Provided acoustic-articulatory evidence for potential “landmarks of resistance” to a complete disguise of “hallå”, namely:

(1) Consistent quasi-uniformity of vocal-tract shapes through the landmarks of /l/

(2) Short sequence of landmarks of coarticulation resistance fro and to /l/

12

Ways Forward IN THE NEAR-TO-MEDIUM TERM:

1) To proceed with the other situational varieties of “hallå” 2) To strengthen the viability of the approaches presented here

IN THE LONG TERM: 1) To acquire further data from the same imitator2) To acquire “hallå” data from native speakers of Swedish dialects3) To enrol more imitators – both professional & naïve

13

Tack så mycket(thank you very much)

3D sequence of Vocal-Tract Shapes through 45 landmarks of “hallå”