@techreport{oai:ipsj.ixsq.nii.ac.jp:00080409, author = {福田, 隆 and 市川, 治 and 西村, 雅史 and Takashi, Fukuda and Osamu, Ichikawa and Masafumi, Nishimura}, issue = {21}, month = {Jan}, note = {近年,特徴空間上の識別学習 (fMMI) が注目され,多くの認識システムで効果を挙げている.通常,識別的特徴変換器には,MFCC+ 動的特徴やセグメント特徴量 +LDA などのスペクトル変動情報を含む特徴パラメータが入力され,それが正準化空間に写像される.特徴空間上の識別学習は近代音声認識において不可欠な要素であるが,低 SNR 環境ではまだ改善の余地がある.本報告では,識別的特徴変換の枠組みに,雑音環境で頑健な性質を示す長時間スペクトル変動情報を組み込むことを提案する.提案手法は低 SNR 環境下で MFCC と動的特徴からなる標準的な特徴ベクトルセットと比較して 6.3% の性能改善を達成した., Discriminative training of feature space using a maximum mutual information (fMMI) objective function has been shown to yield remarkable accuracy improvements. MFCC and dynamic features or LDA features are usually used for discriminative feature transform to map the features into canonicalized feature space. Discriminatively trained feature space transforms are essential for modern speech recognition but still need further improvement for low SNR conditions. In this paper, we show how noise-robust long-term temporal features can be combined with fMMI to build better discriminative models for noisy speech. The fMMI combined with long-term temporal features achieved 6.3% error reduction on average in low SNR environments when compared to the short-term temporal features alone.}, title = {特徴空間における長時間スペクトル変動成分の識別学習}, year = {2012} }