破裂部スペクトルによる日本語無声破裂子音の識別
Discrimination of Japanese voiceless stop consonants by burst spectrum

Shigeyoshi Kitazawa, Shuji Doshita
1984 Acoustical Science and Technology  
We determined the extent t。 which the burst spect ・ um ・ f Japan ・ ・e P . t , k c・ uld be correctly classi 丘ed by a set of voWel and speaker indeperident linear ' discriminant functlons . The stops were produc ε d i血 the gontext of eV utterarices by 28 male speakers . Following the invariance hypothesis of gross shape of burst spectrum , the 丑 ormalized power spect τum was computed fro 皿 the visually extracted burst portion by the autocorrelation 皿 ethod of Iinear prediction analysis . The
more » ... analysis . The spectrum ノ was represented with 28 ・ dimensional components corresponding nearly to critical bandwidth filter outputs 叩 to 9. 25 kHz . Quadratic transformatien of some pairs of components was useful for、 classi 丘cation . Using rather small nu 皿 ber of components selected by stepwise process , the linear functions discriminated conso ・ nants correctly more than 80 % , and about 90 % if the target vpwel was knewn .
doi:10.20697/jasj.40.5_332 fatcat:6ic2wkuiuvdhvcdwgszksk4hti