PRODUCTION AND PERCEPTUAL ANALYSIS OF SPEECH PRODUCED IN NOISE

When exposed to noise, speakers modify the way they speak, possibly in an effort to
maintain intelligible communication. These modifications are collectively referred to
as the Lombard effect. The work described in this thesis compares speech production
changes induced by noise with various spectral and temporal characteristics, and
explores the perceptual consequence of these changes. The thesis consists of a series
of experimental studies, which involve the analysis of speech corpora collected
under different noise conditions, with and without a communicative task.
Intelligibility is also measured and predicted using a computer model.
The first study concerns the acoustic and phonetic consequences of N-talker
“babble” noise on sentence production for a range of values of N from 1 (competing
talker) to “infinity” (speech-shaped noise). The effect of noise on speech production
increased with N and noise level, both of which act to increase the energetic masking
effect of the noise. In a background of stationary noise, noise-induced speech was
always more intelligible than speech produced in quiet, and the gain in intelligibility
increased with N and noise level, suggesting that talkers modify their productions to
ameliorate energetic masking at the ears of the listener.
The effect of low- and high-pass filtered noise on speech production was also
examined to address the issue of whether speakers can compensate for energetic
masking by actively shifting their spectral energy to regions least affected by the
noise. Little evidence was found that speakers are able to modify their speech
production to take advantage of those spectral regions clear of noise.
To evaluate the origin of the increased intelligibility of Lombard speech, the
fundamental frequency and spectral tilt of speech produced in quiet were artificially
manipulated to match those of speech produced in speech-shaped noise. A perceptual
evaluation showed that spectral flattening made a larger contribution to Lombard
speech intelligibility, but failed to find an influence of an increase in fundamental
frequency. A computational modeling study indicated that durational changes could
also play an important role in increasing intelligibility. These findings suggest that
speech modifications which reallocate energy in time and frequency to introduce more
“glimpses” of clean speech in the presence of noise are able to contribute to speech
intelligibility.
An analysis of the effect of noise on speech production requires material recorded
while undertaking realistic tasks. The effect of a communication factor was explored
using conversational speech collected in the presence of maskers with differing
degrees of energetic and informational masking potential. The size of speech
production changes was found to scale with the energetic masking potential of
background noise, extending the findings with read speech to a communicative task.
In addition, relative to the non-communicative task, speakers exploited temporal
planning to reduce the amount of overlap with a modulated background noise, an
effect which was stronger when the noise contained intelligible speech.
In conclusion, the strategies used by talkers to promote successful speech
communication under various noise conditions reported in this thesis could enable
spoken output applications such as dialogue systems to adapt to communicational
environment.