ACOUSTIC MODEL REGULATOR AND PROGRAM

• Main Author: HARADA MASAHARU
• Format: Patent
• Language: English; Japanese
• Published: 14.12.2015
• Subjects: ACOUSTICS; MUSICAL INSTRUMENTS; PHYSICS; SPEECH ANALYSIS OR SYNTHESIS; SPEECH OR AUDIO CODING OR DECODING; SPEECH OR VOICE PROCESSING; SPEECH RECOGNITION

PROBLEM TO BE SOLVED: To improve voice recognition performance.SOLUTION: A learning section (12) learns the parameter of an initial hidden Markov model that includes a plurality of states arranged in line in a time axis direction and represents each of phonemes using voice data for learning on which a vocalization label corresponding to the kind of the phoneme is attached. An acquisition section (14) performs voice recognition of the voice data for learning using an acoustic model represented by the hidden Markov model learned by the learning section to acquire the duration length of each phoneme. An adjustment section (16) performs adjustment so as to increase the number of states included in the hidden Markov model representing the phoneme belonging to a kind whose representative value of the duration length obtained for each kind of the phoneme is a first specified value or larger. 【課題】音声認識性能を向上させる。【解決手段】学習部（１２）は、時間軸方向に並んだ複数の状態を含み、かつ音素の各々を表す初期隠れマルコフモデルのパラメータを、音素の種類に対応する発声ラベルが付された学習用音声データを用いて、学習する。取得部（１４）は、学習部で学習された隠れマルコフモデルで表された音響モデルを用いて学習用音声データを音声認識することにより、各音素の継続時間長を取得する。調整部（１６）は、音素の種類毎に求めた継続時間長の代表値が第１所定値以上の種類に属する音素を表す隠れマルコフモデルに含まれる状態の数を増加するように調整する。【選択図】図１