Digital Signal Processing Reference
In-Depth Information
attention to the traffic and the driving instead of searching knobs and looking
around in the cockpit at high speed.
The state of the art, however, looks somewhat different. There are very
few speech control solutions that can meet the high requirements of an
acoustically difficult car environment.
Unsatisfactory functionality is still the major reason for the comparably
poor customer's acceptance of speech control. One could improve the
acceptance with high level speech recognizers and sophisticated user
interfaces. On the other hand, speech control in cars is expected to provide a
high accuracy, to be robust against noise and environmental variations, but
also to be very cost efficient. Typical requirements for speech control
systems in particular apply to cars:
High recognition accuracy,
Use of command words or command phrases,
Fixed set of speaker-independent commands, and
Programmable set of speaker-dependent commands.
Furthermore, some special requirements for in-car applications can be
defined:
Speed-independent background noise characteristics,
Low-cost embedded solution,
'Push to talk' to prevent false alarms,
Known (mostly stationary) speaker-microphone distance.
In the following section, a novel approach to command word recognition
is presented and set into relation to established techniques such as HMM or
DTW. The following section shows, how a hands-free solution is
implemented based on this core algorithm. Finally, measures are presented to
improve the module's robustness against typical background noises. The
chapter is finished with a discussion on future developments.
2.
RECOGNIZER TECHNIQUES
2.1
Basic approaches for speech recognition
Hidden Markov Models (HMM) are the state of the art technology and
widely used in practical applications. In isolated word recognition, a HMM
