Digital Signal Processing Reference
In-Depth Information
These examples demonstrate three types of interaction between two
speakers. For the cases in which more than two speakers are involved, the
same approach is applied to check the interaction type and goal completeness,
and generate the response to the speakers.
4.
EXPERIMENTAL RESULTS
To test whether the proposed methods were feasible, an experimental
environment was set up in a 1,600 CC automobile. The recording device was
a notebook computer together with a PCMCIA multi-channel recording card
and four omni-directional microphones. We developed an MSDS which was
capable of answering user queries in three application domains, i.e., route
guidance, weather forecasting, and stock prices. A GPS (global positioning
system) receiver was mounted on the car to acquire its current position. The
information about weather forecasts and stock prices was stored in a remote
server that was able to get up-to-date information through the Internet. When
inquires about these domains were made, a short message was issued from an
embedded cell phone to the server. After a database query on the server-end,
the query result, in short message format, was sent back to the cell phone and
was interpreted, resulting in the desired information.
Thirty-two speakers aged 17 to 35 participated in our experiment. Before
the experiment, testers were briefly informed regarding the capability and
domains of the system. Two types of experiment were carried out in our work.
The first was for active speaker determination because knowing the active
speaker is essential in order for the dialogue manager to make a correct
response. The second was the evaluation of the proposed MSDS; both
subjective and objective evaluation metrics are reported in our experiments.
4.1
Experimental Results of Active Speaker
Determination
For the experiment of active speaker determination, we set up four
different configurations of microphone placement, as shown in Figure 4-4.
Four speakers, and were in the upper-left, upper-right, lower-left,
and lower-right corner of the car, respectively, as indicated in Configuration 1
of Figure 4-4.
