NOISE COMPENSATION IN SPEAKER-ADAPTIVE SYSTEMS

• Main Authors: JAVIER LATORRE-MARTINEZ, YANAGISAWA KAYOKO, VINCENT PING LEUNG WAN
• Format: Patent
• Language: English; Japanese
• Published: 10.08.2017
• 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 adapt a related acoustic model to speech vectors in a text-to-speech system.SOLUTION: A speech including noise is received S301. Until a set of speaker transforms most matched with the input speech is determined, a speaker transform of an acoustic model is corrected S303. Features relating to characteristics of the noise are determined from the input speech S305. The speaker transform calculated in S303 and a noise parameter calculated in S305 are utilized for determining a speaker transform of "no noise", namely, a speaker transform that may be obtained in the case where the received speech is obtained under a condition of no noise.SELECTED DRAWING: Figure 5 【課題】テキストツースピーチ（text-to-speech）のシステムにおいて発話ベクトルに関連付ける音響モデルを適応させる。【解決手段】雑音を含む発話が受け取られるＳ３０１。入力発話と最もよく一致する話者変換のセットが決定されるまで、音響モデルの話者変換が修正されるＳ３０３。雑音の特性に関係する特徴が入力発話から決定されるＳ３０５。Ｓ３０３において計算された話者変換及びＳ３０５において計算された雑音パラメータが、「無雑音」の話者変換、すなわち、受け取られた発話が、無雑音の条件のもとで得られた場合に得られていたであろう話者変換を決定するために利用される。【選択図】図５