Digital Signal Processing Reference
In-Depth Information
take?” “Is it a right or left turn?” “How do I get to the airport?” “I'm lost.
Help me.”
Digit Sequences
:
each speaker produced 16 digit strings from a
randomized 75 digital string set. Examples include: telephone numbers (425-
952-54o0), random credit card numbers (1234-5621-1253-5981), and
individual numbers (0,o,#86, *551).
Say and Spell Addresses:
a randomized set of 20 strings of
words/addresses were produced, with street names spelled. Some street names
are used for all cities, some were drawn from local city maps. Examples
include: Park Place, Ivy Circle, 3215 Marine Street, 902 Olympic Boulevard.
Phonetically Balanced Sentences
: each speaker produced a collection of
between 20-30 phonetically balanced from a set of 2500 sentences [prompts
are randomized]. Examples include: “This was easy for us.” “Jane may earn
more money by working hard.”
Dialog Wizard - of - Oz Collection
: each speaker from the field called an
on-line navigation system at CSLR, where a human wizard-of-oz like (WOZ)
operator would guide the caller through three different navigation routes
determined for that city. More than 100 potential destinations were previously
established for each city between the driver and WOZ human operator, where
detailed route information was stored for the operator to refer to while the
caller was on the in-vehicle cell-phone. The list of establishments for that
city were points of interest, restaurants, major intersections, etc. (e.g., “How
do I get to the closest police station?”, “How do I get to the Hello Deli?”).
The user calls using a modified cell-phone in the car, that allows for data
collection using one of the digital channels from our recorder. The dialog was
also recorded at CSLR where the WOZ operator was interacting with the field
subject.
The 500 speaker corpus was fully transcribed, labeled, spell checked,
beamformed/processed and organized for distribution. The un-processed
version contains well over 600GB of data, and the processed version consists
of a hard-disk release of approximately 200GB. Figure 2-3 shows the age
distribution of the CU-Move corpus (further details presented in [26,27]).
