| 摘要 |
PROBLEM TO BE SOLVED: To provide a method, an apparatus, and a program for phoneme determination that can accurately discriminate a vocal sound and accurately specify the section of the vocal section, i.e. vocal sound borders. SOLUTION: In the phoneme determining method, an input speech signal is divided into a plurality of bands, frame by frame, through mel frequency division (S1); power of each band is found and an acoustic feature quantity vector of each frame is generated (S2); and acoustic feature quantity vectors of those kinds are used to generate an HMM (Hidden Markov Model) as to respective vocal sounds or vocal sound boundaries. Computation is so carried out (S3) that a series of HMMs corresponding to previously know vocal sounds or vocal sound boundaries of the input speech signal, and a previously found feature quantity vector sequence and likelihood become maximum; and information (label) representing the vocal sounds or vocal sound boundaries is imparted to respective frames of the speech signal at this time (S4). COPYRIGHT: (C)2004,JPO
|
