Original scientific paper
Similarity and dissimilarity indices for the speakers of the Croatian language in filtered and non–filtered conditions
Gordana Varošanec–Škarić
orcid.org/0000-0002-0864-4252
; Filozofski fakultet, Zagreb
Iva Pavić
; Filozofski fakultet, Zagreb
Gabrijela Kišiček
; Filozofski fakultet, Zagreb
Abstract
The main goal of this research was to compare the acoustic–statistical measurements of
similarity index (R) and dissimilarity index (SDDD) on the basis of long term average spectra
(LTASS). The collected speech samples consisted of both spontaneous speech and reading of a
standardized text which were recorded in controlled conditions. The recorded speakers (N =
86) originated from 8 larger and dialectically different Croatian towns. The cities chosen for
this research represented Sthokavian dialect (Zadar, [ibenik, Dubrovnik), Chakavian dialect
(Rijeka, Pula), Kaikavian dialect (Zagreb, ^akovec) and Chakavian–Sthokavian variety (Split).
Speech samples (duration 60 s) were edited for each speaker in both conditions (filtered and
non–filtered speech) and compared on the basis of LTASS – non filtered conditions (0 – 10
kHz) and filtered (0.8 – 4 kHz). Using index R and index SDDD intraspeaker variations and
interspeaker variations were compared respectively for male and female speakers. Results
of intraspeaker variations showed that average values of similarity index (R) in non–filtered
conditions were between 0.94 for male speakers in reading texts to 0.98 for female speakers in
reading text and spontaneous speech. Results of interspeaker variations showed lower values of
index R in the non–filtered conditions: from 0.86 in spontaneous speech to 0.94 in reading text
for female speakers. Average values of R in filtered conditions for intraspeaker variations were
between 0.83 for both female and male speakers in spontaneous speech to 0.95 in reading texts.
Average values of R index in filtered conditions for interspeaker variations were significantly
lower; from 0.57 for male spontaneous speech to 0.9 for female reading texts. Average values of
index SDDD in non–filtered conditions for intraspeaker variations were generally lower – from
2.27 for female speakers to 3 for male speakers in reading. SDDD index showed higher values
in non–filtered conditions for interspeaker variations; from 4.75 in female reading speech and
male spontaneous speech to 5.12 for male reading speech. In filtered conditions intraspeaker
variations resulted with SDDD index between 2.14 for male reading speech to 3.01 for female
spontaneous speech. As expected, results in filtered conditions for interspeaker variations
showed higher values of SDDD index, from 3.06 for female to 4.71 for male reading speech. The
differences between similarity index (R) in intraspeaker variations were statistically significant
for female speakers (p < 0.0001) and for male speakers (p < 0.05) in both spontaneous speech
and reading. Results of interspeaker variations showed statistically significant differences in
similarity index (R) for male speakers (p < 0.0001 in reading and p < 0.0001 in spontaneous
speech) and female speakers (p < 0.0001 in reading and p < 0.0001 in spontaneous speech) and
statistically significant dissimilarity index (SDDD) differences for male speakers (p < 0.0001 in
reading and p < 0.0001 in spontaneous speech) and female speakers (p<0.0001 in reading and
p<0.0001 in spontaneous speech). Overall results of this research show that acoustic–statistical
measurement of similarity and dissimilarity indices are a useful method in speaker recognition
in forensic phonetic expertise. Further on, results show that speaking conditions should not be
neglected in forensic phonetic cases.
Keywords
speaker identification; similarity index (R); dissimilarity index (SDDD); long term average spectra (LTASS); filtered speech; non–filtered speech; spontaneous speech; reading speech
Hrčak ID:
125115
URI
Publication date:
16.7.2014.
Visits: 2.051 *