Automatic generation of simple Markov model stunted baseforms for words in a vocabulary

DETAILED DESCRIPTION It is, therefore, an object of the present invention to provide a baseform which significantly avoids the disadvantages of the fenemic baseform and phonetic baseform while, at the same time, including major benefits of each.
Moreover, it is an object of the invention to automatically form short simplified phonetic-type Markov model word baseforms.
In accordance with the invention, it is presumed that each word in the vocabulary is defined by an existing fenemic baseform.
It is also presumed that an alphabet of composite phones has been derived.
Preferably the composite phones are chosen to be similar to elements in a phoneme alphabet.
In this preferred case, each phone machine corresponds to an element of the phoneme alphabet and stores transition and label probabilities that are used in matching the phoneme element to the fenemes in a string of fenemes.
There may be less elements in the composite phone alphabet than in the fenemic phone alphabet.
The simplified baseform achieved by the present invention is constructed of composite phones in the following way.
A subject word is uttered once or, preferably, a plurality of times.
With each utterance a string of fenemes is generated.
Based on the string or strings of fenemes generated for the subject word, a fenemic word baseform having the highest (joint) probability of producing the string(s) is selected.
Each string for the subject word is then aligned against the fenemic baseform.
In this way, each fenemic phone in the fenemic baseform is aligned against corresponding fenemes in each string (for the subject word).
Each fenemic phone in the fenemic word baseform is replaced by the composite phone corresponding thereto, thereby forming a composite phone baseform.
The following steps are then performed: 1.
A pair of adjacent composite phones is selected.
2.
For each string of fenemes, the substring of fenemes that is aligned against the selected pair of adjacent composite phones is determined.
3.
The one composite phone having the highest joint probability of producing the respective determined substrings for all the strings of fenemes is determined