Robotics Reference
In-Depth Information
alone, without the need to add any significant linguistic or other knowl-
edge. When the corpora are large enough, applying inductive learning
and neural networks to them, will enable systems to be developed in
which the rules are derived automatically from the corpora.
During the last decade of the twentieth century, advances in NLP
were becoming more rapid and more impressive, partly because of an
increasing availability of large corpora in electronic form, partly due to
researchers having better access to faster computers with bigger memo-
ries, and partly due to the immense dynamic and latent powers of the
Internet. Statistical approaches were found to be successful in solving
many generic problems in computational linguistics, for example part-
of-speech tagging and disambiguation, and have come into standard use
throughout the field of NLP. I believe that this trend towards the greater
use of techniques based on large corpora will flourish, and that as the
corpora get bigger, so conversational and translation programs will im-
prove. A three-year old child does not need to parse a sentence in order
to know how to respond to something said by a parent or, if the child is
bilingual, to translate from one of its languages to another, so why should
a robot? Just as the domain of Chess was conquered by programs that
think about the game in quite a different way to the thinking of human
grandmasters, so the conversational and translation programs of future
decades will perform their tasks in ways that are very different from the
thought processes of humans.
Text-to-Speech Synthesis
One of the key elements in the communication process between humans
is speech, and for robots to be able to communicate with us in a user-
friendly way they must be able to talk like a human being and to say or
read out loud any text, using the same type of intonation and stresses
that we use. The technology for achieving this is called Text-to-Speech
(TTS) and is simply the automatic conversion of written text into spoken
output.
Text-to-speech synthesizers can be divided into three categories: rule-
based (also known as parametric synthesis), articulatory and concatena-
tive. Rule-based synthesizers start with an electronic tone that vibrates
at the same rate (or frequency) as the human vocal chords, and mod-
ifies that frequency continuously—making hundreds of modifications
