In-Vehicle Speech and Noise Corpora - Digital Signal Processing for In-Vehicle Systems and Safety - page 140

of the car during the recordings was 40 mph, and data was collected on a 4-mile
route of concrete roads. The route was selected so as to consist of a combination of
6-lane city roads with higher traffic density and 2-lane concrete community roads
with lower traffic density. The car noise data recording was timed so as to minimize
external traffic noise due to peak hours.
The data collection consisted of two parts. The first set of in-vehicle noise events
were recorded in the University of Texas at Dallas parking lot. For these recordings,
the vehicle was stationary, the windows were closed, and the AC was turned off.
Under these vehicle conditions, the following sound events were collected:
(a) Turn signals (TRN)
(b) Horn (HNK)
(c) Front doors opening and closing (LDR/RDR)
(d) Engine idling (IDL)
(e) Revving (REV)
The average total recording time for these conditions was about 6 min.
The second set of recordings took place on roads around the University of
Texas at Dallas campus. The data was collected only in dry weather conditions,
where the vehicles completed the route twice. The route was 2 mi. long, with
two-to-three lane roads and speed limits ranging from 30 to 40 mph. For this corpus,
the route was divided into seven sections, and a particular noise condition was
assigned to each section.
Figures 9.2 and 9.3 show the designated route. The seven sections of the route are
also shown in the figures. As shown in Fig. 9.2 , two noise conditions (ACWC,
NAWC) were collected in the first loop. During sections A to D of the route, the
windows and AC remained closed. In sections E to G of the route, the windows were
closed and the AC was turned on with blower at full capacity. Meanwhile, Fig. 9.3
shows the four noise conditions recorded during the second loop. Section K of the
route included a speech exercise. Here, the driver was asked to count out aloud from
0 to 9, three times, with the windows closed and AC turned off. Data for NAWC
condition was recorded again in sections L and N. The final recording condition was
ACWC in section H. The average on-the-road recording time was about 21:25 min.
The priorities of this exercise were the NAWC and ACWC conditions as speech
systems encounter these conditions frequently in car environments. Collection setup
was designed to allow for data collection in multiple sessions to ensure that the audio
recording of the car events contained variability due to different road/traffic
conditions. The corpus contains a total of 8 h of car noise data.
9.3 CU-Move
The UTD-VN corpus deals with variability in fixed environments across car makes
and model. Another aspect of in-car acoustics as mentioned in sect. 1 includes
speech variability due to stress along with noise. These are the major causes of
acoustic mismatch in a car environment. The CU-Move corpus focuses on