Digital Signal Processing Reference
In-Depth Information
moment, this costly process is implemented in a distributed evaluation
network, where one computer runs the evolutional engine, which sends the
parameters to optimize over the LAN. All other computers connected to the
network run 1 to 3 recognizer individuals (clients), and send the results of the
evaluation back to the evolution. The training and test data is accessed by
each client over the network. In this way, a speaker independent word
recognition rate of 98.6 % can be reached in the absence of noise (see
Section 4.3).
3.
TARGET APPLICATION AND DIALOGUE
The target application is a telephone control with combined speaker
independent commands for hands-free telephone operation, and speaker
dependent commands for name dialling. The application is fully operable by
voice, because the user is guided through all menus by high quality voice
prompts.
To increase the robustness of the application the recognizer is embedded
in an ergonomic dialog including voice prompts and 'push to talk'. There are
30 speaker independent commands and 30 speaker dependent commands
(plus corresponding actions like a telephone number for each command)
available in the lexicon. Commands are ordered in submenus to enable
functions like user dependent training (storage of new names into the
lexicon), dictation of phone numbers (collection of number chains, repetition
and navigation in the number chain).
The high sound quality of the voice prompts ensure a high acceptance of
the speech control by the user. Prompts are stored in a scalable memory and
are generated by an application-specific word-unit synthesizer. So even large
and well tuned dialogs can be stored with minimum memory requirements.
The dialogue is designed very flat to minimize the distraction of the driver
from traffic.
4.
ROBUSTNESS
Recognition accuracy loss mainly occurs at the speaker microphone path.
Some important influences on the quality of the received speech signal and
methods to overcome them are listed below:
Microphone type.
Electret capacitor microphones with directional
characteristics should be used in cars to attenuate side noises and keep the
system costs low. Microphone arrays can be used for beam forming. At
