Takeshi KAWABATA
Constructive Neural Network for
Speech Recognition
Abstract:A new method to combine phone TDNNs for constructing word or
continuous speech networks is proposed. The method is called
Constructive Neural Network (CNN). The CNN method has the
following important features:
(1) Any word network can be composed of neural phone verifiers
according to its phonetic representation.
(2) Non-linear time warping is realized using only a network
structure including FIR filters which control the duration between
each phone pairs independently.
(3) Word-level backpropagation learning can be used for training
the TDNN phone recognizers.