混合音声からの特定話者信号抽出に関する基礎的考察

概要

論文の詳細を見る
Recently, many researches have been reported on the separation and extraction method to abstract a specific speaker's speech at the conference. This study focuses on the basic method to separate and extract a specific speaker's speech on the premise that there are independent speech signals of male and female speakers in the signals, with a clue of the characteristic for the speech signal in a frequency domain. On a concrete basis, it is the method to extract a specific speakers syllable by (1) discerning the monosyllable that constitutes the objective syllable for extraction from other independent speech signals, (2) defining the average value of its Short Time Fourier Transform as the specific filter of the speaker for the monosyllable and (3) filtering the STFT spectra of the male-female mixed speech through the above-mentioned specific filter of the speaker. The result of a numeric experiment shows that this method can extract the speech in question despite remaining the speech of speakers other than the objective speaker for extraction to an ignorable extent.
秋田工業高等専門学校の論文
2003-02-28