To improve the recognition rate of voice recognition for a new word and to learn a language model during the correction of an original to be used for reading by weighting the frequency of the new word concerning the non- corrected original.
When updating a vocabulary, an operator inputs the latest news original from an input device 120 or disk reader 130. This news original is a source original before correction into original to be read for broadcasting. The inputted original is integrated (stored) in a system memory 11 by a CPU 100 and preserved on a hard disk 140 for long-term integration. The CPU 100 calculates text weight (weight of frequency). Next, the CPU 100 adds the inputted latest news original to the preserved long-term news originals and calculates the frequency of respective words in the long-term news originals. The calculated frequency is preserved on the hard disk 140 for determining the vocabulary to be registered and for calculating the next frequency.
IMAI TORU
ANDO AKIO
Next Patent: VOICE RECOGNITION DEVICE