| 摘要 |
PROBLEM TO BE SOLVED: To reduce the number of acoustic sub-word models as an acoustic model depending on a preceding or following phonemic environment and to apply the model to a recognizing system considerably limited in memory capacity or arithmetic throughput. SOLUTION: Attention is paid to respective vowels ', a, i and a' in phonemes, which constitute 'tadaima (Hello, here I am!)', as an audio language, for example, and when phonemes adjacent to the front and the back of each of concerned vowels are vowels, these vowels are considered but when no phoneme exists at that position, it is handled as silent so that the acoustic sub-word model can be generated. Besides, when phonemes adjacent to the front and the back of the concerned vowel are consonants, a vowel existent next to that constant is considered but when no phoneme exists at such a position, it is handled as soundlessness so that the acoustic sub-word model can be generated. With respect to the audio language of 'tadaima', acoustic sub-word models M1, M2, M3 and M4 are generated.
|
