3. English speech in noise
The stimuli were from a previous study
!
Iverson et al.,
2006
", recorded from a different British English talker !fe-
male
" than used in the rest of this study. The speaker was
recorded saying Say hVd again with 11 words: heed
!/i/", hid
!/(/", head !/!/", had !/a/", heard !///", hud !/#/", hod !/"/",
hard
!/Ä/", hoard !/Å/", hood !/*/", and who’d !/u/". Diph-
thongs were not included because they become unintelligible
when formant movement is removed. Two additional ver-
sions of the vowels were created that:
!1" removed all for-
mant movement, and
!2" equated duration. The changes to
the stimuli were made using Praat
!
Boersma and Weenink,
2005
". Formant movement was removed using linear predic-
tive coding
!LPC" analysis and resynthesis. Specifically, LPC
analyzed the signal from the start of voicing after the /h/ to
the start of the /d/ closure; the signal was inverse filtered to
produce an LPC residual; a time slice of the LPC analysis
was identified that represented the vowel’s target formant
frequencies
!defined as the point where F1 reached a peak";
and this single LPC slice was used to filter the entire LPC
residual. This process created stimuli that retained the natural
F0 of the original stimuli, but had formant frequencies that
remained fixed at each vowel’s target values. Duration was
equated using PSOLA, such that the durations of the /h/, the
/d/ closure, and the vowel were set to the mean values for the
talker.
The speech-shaped noise conformed to CCITT Rec.
G227 and was produced by a Wandel and Goltermann RG-1
noise generator. The signal-to-noise ratio
!SNR" was calcu-
lated for each individual stimulus, by comparing the RMS
amplitude of the stimulus and noise.
Do'stlaringiz bilan baham: |