Digital Signal Processing Reference
In-Depth Information
3.
CU-MOVE CORPUS DEVELOPMENT
As part of the CU-Move system formulation, a two phase data collection
plan was developed. Phase I focused on collecting acoustic noise and probe
speech from a variety of cars and driving conditions. The outcome of Phase I
was to determine the range of noise conditions across vehicles, and select one
vehicle for Phase II collection that is representative of the typical noise
domains experienced while driving. Eight vehicles were used in Phase I
analysis (e.g., compact and two mid-size cars, small and medium pickup
trucks, passenger van, sport utility vehicle (SUV), cargo van). We considered
14 noise conditions in actually driving scenarios. Figure 2-2 summarizes
some of the results obtained from the study, with further details presented in
[26]. The noise level was highest with windows open 2 inches traveling
65mph on the highway, and most quiet when the car was idle at a stop light.
After detailed analysis, we determined the SUV represented the mid-range
noise conditions (noise levels were high for compact cars and low for pickup
trucks).
Next, Phase II speech collection was performed. Since the speaker is
experiencing some level of stress by performing the task of driving the
vehicle, this should be included in the speaker modeling phase. While
Lombard effect can be employed, local state and federal laws in the United
States limit the ability to allow subjects in this data collection to operate the
vehicle and read prompts from a display. We therefore have subjects seated in
the passenger seat, with prompts given on a small flat panel display attached
to the dashboard to encourage subjects to stay focused on the roadway ahead.
Speech data collection was performed across 6 U.S. cities that reflect regional
dialects. These cities were selected to be mid-size cities, in order to increase
the prospects of obtaining subjects who are native to that region. A balance
across gender and age brackets was also maintained. The driver performed a
fixed route similar to what was done for Phase I data collection so that a
complete combination of driving conditions (city, highway, traffic noise, etc.)
was included. The format of the data collection consists of five domains with
four
Structured Text Prompt
sections and one
Wizard-of-Oz
(WOZ) dialog
section:
Navigation Phrases
: collection of phrases useful for In-Vehicle navigation
interaction [prompts are fixed for all speakers]. Examples include: “Where is
the closest gas station?” “How do I get to 1352 Pine Street?” “Which exit do I
