In-Vehicle Speech and Noise Corpora - Digital Signal Processing for In-Vehicle Systems and Safety - page 144

Fig. 9.4 The Scope of UTD-VN vs. CU-Move

corpus is a compilation of events encountered in a car-driving scenario as described
earlier whereas the UTD-VN is a compilation of a few events across various cars
and conditions.

9.4 Noise Analysis and Modeling

The car noise-environment noise samples in both the corpora can be described as a
combination of noise sources active in the car, as well as the acoustic environment
of the vehicle itself. In other words, the resultant car noise is a function of car-
independent noise nðÞ

and car-dependent noise n c ðÞ . Here, an additive model for

n c ðÞ is assumed. This is illustrated in Fig. 9.5 . Depending on the relative domi-

nance of the constituent noises, the overall resultant noise observed can be of three
primary types.

^

• Car Internal Dominant Noise : If car-dependent sounds such as air conditioning,
horn, and engine sounds dominate, then the resulting noise n e is unique to the
specific car producing the sound (i.e., if n e <<

). For
purposes of car verification/platform identification, this forms the most condu-
cive scenario. For speech systems, it means that car specific models might be
optimal for the best performance in specific car environments.
• Car-Environment Dominant Noise : If the observed sound is the sound of the car
interacting with its environment, such as the sound of wheels on the road or wind

n ce then n c ðÞ ^ n c ðÞ