Telugu  

and  Japanese:  a  case  for  

the  use  of  rhythmic  similarities  
in  speech  segmentation

Dr. Lalita Murty

University of York
Segmentation  Processes
—  Speech   is   continuous   and   the   only   way  
to   understand   an   u;erance   is   to  
understand   the   words   of   which   it   is  
made  up.
—  There   is   nothing   in   the   speech   signal   to  
mark  the  boundaries  of  words.  
—  T h e r e f o r e ,   a   p r e r e q u i s i t e   f o r    
u n d e r s t a n d i n g   w o r d s   i s   t h e  
segmentation   of   individual   words   that  
make  up  the  speech  signal.
—  Words,  however,  cannot  be  recognised  in  
a   linear   fashion   one   by   one,   primarily  
because   words   overwhelmingly   contain  
other  words.  E,g.  star  may  turn  out  to  be  
startle,  start,  starch,  stark?    
Rhythm  in  Listening

Listeners  process  spoken  language  in  

ways  that  are  subtly  suited  to  the  
phonological  structure  of  their  native  
languages.

One  aspect  of  language  structure  which  
affects  listening  processes  is  Rhythm.
Linguists  have  identified  3  types  of  rhythmic  
pa;erns  in  languages:  
—  Stressed  timed  languages  like  English  and  
Dutch
—  Syllable  timed  languages  such  as  Spanish  
and  French.
—  Mora  based  languages  like  Japanese  and  
some  languages  spoken  in  South  India.
Pycholinguistic  research  has  shown  that  
similar  rhythmic  structure  encourage  
similar  types  of  processing  in  listening.
Languages  that  differ  in  rhythmic  structure  
encourage  different  types  of  processing
—  French  listeners  and  not  English,  use  
the  syllable  as  a  segmentation  cue  
(Cutler,  Mehler,  Norris  and  Segui,  
1983,  1986)
—   English  listeners    have  a  stress-­‐‑based  
segmentation  strategy  (Bu;erfield  and  
Cutler,  1988;  Cutler  and  Norris,  1988;  
Cutler,  Mehler,  Norris,  and  Segui,  
1989,  1992).
—  French  and  English  have  been  held  as  
the  examples  supporting  the  rhythmic  
distinction  between  syllable-­‐‑timed  
and  stress-­‐‑times  languages.
Speech  segmentation  in  Japanese
70 tanishi
—  Japanese-­‐‑speakers   tanshi
detect  word-­‐‑ 60
initial  targets  such  
as  ta-­‐‑,  tan-­‐‑  if  the   50
target  is  exactly   40
one  mora  (ta-­‐‑  in   %
tanishi  or  tanshi)   missed 30
or  two  (tan-­‐‑  in   20
tanshi)
10
—  They  fail  to  detect     0
targets  not   ta- tan-
aligned  with  mora  
boundaries    
(tan-­‐‑  in  tanishi)
(Otake,  Hatano,  Cutler  &  Mehler,  1993)
Segmentation  of  non-­‐‑native  speech

Predictions:  
—  1.  Listening  to  a  non-­‐‑native  language  with  the  same  
rhythmic  structure  should  be  easier  than  listening  
to  a  non-­‐‑native  language  with  a  differing  rhythmic  
structure.

—  2.  “Native”-­‐‑like  segmentation  of  a  given  language  

by      non-­‐‑native  listeners  should  only  be  possible  for  
listeners  whose  native  language  encourages  the  
same  segmentation  procedure.    Listeners  with  
other  languages  will  apply  their  different  
procedures,  inefficiently.
The  present  study
Aim:  
The   study   investigates   the   possibility   of  
rhythmic   similarity   classes   comprising   of  
sets   of   potentially   unrelated   languages,  
Japanese   and   Telugu   both   of   which   are  
considered   to   be   MORAIC   languages  
(Caldwell,  1856;  Shiba,  1972;  Ohno,  1981).
The   prediction   is   that   in   a   fragment  
monitoring  task,  listeners  of  both  languages  
will  use  the  mora  as  a  cue  for  segmentation  
in  their  respective  languages  as  well  as  each  
others’  languages.
What  is  a  mora?
A  mora  is  a  subsyllabic  unit.  
Vowels  may    be  short  (giving  the  syllable  
structures  V,  CV.  CCV)  or  long  (VV,  CVV,  
CCVV);
 a  consonantal  coda  can  only  be  a  Nasal  
(VN,  CVN,  CCVN)  or  a  geminate  with  
the  onset  of  the  following  syllable    (thus  
only  Nasals  can  be  word  final  codas).  
Syllables  with  a  short  vowel  consist  of  one  
mora,  syllables  with  a  long  vowel  or  a  
nasal  coda  consist  of  two  moras  
(bimoraic).  
Japanese  and  Telugu  material
Material:  16  meaningful  Japanese  and  Telugu  words  
in  eight  pairs  were  respectively  presented  :  
Japanese:  tanshi-­‐‑tanishi;  monaka-­‐‑  monka;  kinori-­‐‑
kinri.
Telugu:  penku-­‐‑penamu;  winta-­‐‑winati;  manawi-­‐‑
manta;  chembu-­‐‑chemata
(Murty,    Cutler  and  Otake:  2007)
These  16  pairs  were  embedded  in  250  other  words  
arranged  in  64  sequences.  Each  target  word  
appeared  twice:  once  in  the  one  of  the  first  32  
sequences  and  once  in  the  of  the  last  32  sequences.  
The  sequence  length  varied  form  3-­‐‑6  words.
Fragment  or  phoneme  monitoring  task

Subjects   were   asked   to   listen   for   a   specified  

fragment   which   would   occur   at   the   beginning   of  
one  word  in  a  list  of  words  .  E,g  they  listened  for  
‘che’   in   a   list   of   5-­‐‑7   words   which   includes   the  
word   ‘chemata’   and   pressed   a   bu;on   as   soon   as  
they  heard  it
Subject  listened  to  2  types  of  words  and  
pressed  a  bu;on  when  they  heard  either  
one    of  two  types  of  fragments:
Word  type:  CVNCV;  CVCVCV
Target  Types:  CV;  CVN
E.g.:
 ta  (CV)  in  tanshi  (CVNCV)  or  in  tanishi  
(CVCVCV)
Tan  (CVN)  in  tanshi  (CVNCV)  or  in  tanishi  
(CVCVCV)  
Prediction
—  Subjects  will  find  it  easy  to  identify  ta  
in  tanshi  (CVNCV)  or  tanishi  
(CVCVCV)  because  it  is  equal  to  one  
mora  and  syllable
—  Subjects  will  find  it  easy  to  identify  
tan  in  tanshi  because  it  is  equal  to  one  
syllable  (tan-­‐‑)  and  2  moras  –  ta-­‐‑  and  n  

—  Subjects  will  have  problems  
identifying  tan  in  tanishi  because  it  is  
more  than  one  syllable  and  less  than  a  
mora.
 Japanese  and  Telugu  rhythm
•  Telugu  has  mora-­‐‑based  rhythm,  like  Japanese
•  Telugu  listeners  show  a  Japanese-­‐‑like  pa;ern  with  input  in  
Japanese;  Japanese  listeners  show  the  same  pa;ern  with  input  
in  Telugu.
%  Miss  Rates  

Telugu-Japanese Japanese-Telugu
40
40

30 30
manishi
tanishi
mandu
20 tanshi
20

10 10
ta- tan- ma- man-

(Murty,  Otake  &  Cutler,  2007)

Overall  findings  of  the  study
Japanese  Materials
—  Japanese-­‐‑speakers  detect  word-­‐‑initial  targets  such  as  ta-­‐‑,  tan-­‐‑  
if  the  target  is  exactly  one  mora  in  both  Japanese  and  in  
Telugu    (ta-­‐‑  in  tanishi  or  tanshi;  ma-­‐‑  in  manishi  or  mandu  )  or  
two  (tan-­‐‑  in  tanshi;  man-­‐‑  in  mandu)
—  They  fail  to  detect    targets  not  aligned  with  mora  boundaries    
(tan-­‐‑  in  tanishi;  man-­‐‑  in  manishi)  
—  Telugu  listeners  ‘  performance  on  Japanese  words  was  very  
similar  to  that  of  native  Japanese  speakers,  the  miss  rate  for  
CVN  target  in  CVNCV  words  was  double  that  in  the  other  
three  conditions.  And  the  responses  to  those  targets  are  
slower  than  in  any  other  three  conditions.
Telugu  Materials
—  Telugu  speakers  performance  on  Telugu  words  was  mixed  as  
there  were  no  significant  results,  but  the  general  advantage  
for  CV  targets  in  both  CVCVCV  and  CVNCV  words  was  
similar  to  that  of  Japanese  speakers.  
—  Japanese  listeners  responded  to  Telugu  
material  in  the  same  way  as  they  did  to  in  
words  their  own  language.
—  Telugu  speakers  clearly  used  strategies  
similar  to  those  of  Japanese  speakers  when  
listening  to  Japanese  words,  thus  showing  a  
tendency  to  use  mora-­‐‑based  listening  
strategies,  even  if  they  did  not  appear  to  be  
as  strongly  dependent  as  their  Japanese  
counterparts  on  such  strategies  when  
listening  to  words  in  their  native  language.
—  Nevertheless,  the  results  on  the  whole  lend  
support  to  the  rhythmic  segmentation  
hypothesis.