INTENDED SOUND EXTRACTION DEVICE AND METHOD FOR EXTRACTING INTENDED SOUND

PROBLEM TO BE SOLVED: To determine the presence of silence by frequency and accurately detect a silent section, thereby simplifying a system.SOLUTION: An intended sound extraction device operates as follows: converting multiple types of digital signals into frequency-domain signals formed of a frame that includes a plurality of spectra specified by a frequency bin; detecting a silent section including no sound, on the basis of the multiple types of frequency-domain signals thus obtained; estimating noise on the basis of the detection signal; reducing, for each spectrum of the frequency bin, noise in one type of frequency-domain signal, out of the multiple types of frequency-domain signals, on the basis of the estimation signal; and converting the one type of noise-suppressed frequency-domain signal into a time-domain signal. The silent section is detected on the basis of a phase difference between the spectra of frequency bins of the same number in frames of the same number of the multiple types of frequency-domain signals.SELECTED DRAWING: Figure 2 【課題】周波数毎に無音声の有無が判断でき正確な無音声区間の検知ができるようにしてシステムの簡素化を実現する。【解決手段】複数系統のデジタル信号を周波数ビンで特定されるスペクトルが複数含まれるフレームで構成された周波数領域の信号に変換し、得られた複数系統の周波数領域の信号に基づいて音声が存在しない無音声区間を検知し、検知信号に基づき雑音を推定し、推定信号に基づいて複数系統の周波数領域の信号の内の１つの系統の周波数領域の信号に対して周波数ビンのスペクトル毎に雑音を抑圧し、雑音が抑圧された前記１つの系統の周波数領域の信号を時間領域の信号に変換する。無音声区間の検知を、複数系統の周波数領域の信号の同一番号のフレームの同一番号の周波数ビンのスペクトルの位相差に基づいて行う。【選択図】図２