TITLE COMPUTER COMMANDS BY VOICE WITH A FEATURE- BASED TECHNIQUE. AUTHOR JITDAMRONG PREECHARSUK DEGREE MASTER OF SCIENCE PROGRAMME IN COMPUTER SCIENCE FACULTY FACULTY OF SCIENCE ADVISOR DAMRAS WONGSAWANG CO-ADVISOR CHOMTIP PORNPANOMCHAI

movedearΤεχνίτη Νοημοσύνη και Ρομποτική

17 Νοε 2013 (πριν από 3 χρόνια και 11 μήνες)

87 εμφανίσεις


TITLE COMPUTER COMMANDS BY VOICE WITH A FEATURE- BASED
TECHNIQUE.
AUTHOR JITDAMRONG PREECHARSUK
DEGREE MASTER OF SCIENCE PROGRAMME IN COMPUTER SCIENCE
FACULTY FACULTY OF SCIENCE
ADVISOR DAMRAS WONGSAWANG
CO-ADVISOR CHOMTIP PORNPANOMCHAI

ABSTRACT This research studies the solution to a computer interface problem by using
human voice instead of input devices. Currently, the computer system is
controlled through input devices such as keyboards, a mouse and other
devices. This research proposes that the computer be commanded by voice.
The basic command groups in English were used in this research and these
command groups were isolated-words. The Computer Commands by Voice
(CCV) prototype was developed based on a speaker-dependent approach
using a feature-
b
ased technique. The important features used in the CCV
were syllables and Mel-Frequency Cepstrum Coefficient (MFCC) features.
The CCV system consisted of 2 important parts. The first part was training,
which identified the start and end points of each command and then extracted
the important features. The features were stored as reference template files.
The second part was recognizing similar to the training part this part also
includes the recognition subpart. The recognition subpart recognizes the
commands stored in the reference template files groups and follows the
user’s commands that are recognition result. The CCV system was tested
using 10 commands. The commands were classified into 3 groups; one, two
and three-syllable groups. Each command was trained 10 times. The system
consisted of 100 command templates and used a sample rate of speech signal
at 8 KHz. The experiment was conducted under 2 environments; official
environment and quiet environment. In both environments, the percentage o
f

correctness was 90% and 95%, respectively conducted by the same subject.
However, the correctness was 36% and 38% respectively, conducted in both
environments by 2 subjects (trained by one and recognized by another) of the
same gender. The results were 28% and 31%, respectively when recognized
and trained by 2 persons of the opposite gender. It was found that the
experiments yield satisfactory results when the person who trains and
recognizes is the same and the commands have been trained before.
Furthermore, the experiment was conducted with untrained commands in
one, two and three-syllable groups. The untrained commands were
commanded 10 times each. The percentage of correctness was found to be
93% and 97% under the official and quiet environment, respectively
conducted (trained and recognized) by the same person. However, the
correctness was 98% and 100%, respectively when trained and recognized by
2 persons of the same gender while it was 90% and 94%, respectively when
trained and recognized by 2 persons of opposite gender. The indications from
this study are that the CCV prototype is able to recognized untrained
commands while efficiently recognizing trained commands.
KEYWORD VOICE RECOGNITION / SPEECH RECOGNITION / PATTERN
RECOGNITION / FEATURE EXTRACTION / FEATURE-BASED /
VOICE COMMANDS