一、特有名詞
acoustic scale: 通常設置爲0.1,意味着聲學模型的log概率得到的權重遠低於語言模型。在Score的腳本中,經常會看到一系列要搜索的語言模型權重(例如7到15的範圍)。 這些可以解釋爲聲acoustic scale的倒數; 這是兩者對維特比解碼很重要的比例
alignment:表示一句話通過維特比(最佳路徑)對齊的HMM狀態序列表示,與transtions-id同義。show-alignment顯示
likelihood:使用標準的交叉熵訓練的神經網絡系統,我們通過將對數概率除以上下文相關狀態的先驗來獲得“僞似然性”。
posterior:後驗概率,Alignments and lattices can be converted to posteriors over transition-ids (see lattice-to-post.cc), or over lattice arcs (see ali-to-post.cc and lattice-arc-post.cc). Posteriors over transition-ids can be converted to posteriors over pdf-ids or over phones; see the tools ali-to-post.cc, post-to-pdf-post.cc and post-to-phone-post.cc
二、關於lattice的工具
latgen-faster-mapped: 生成lattice
lattice-scale:Apply scaling to lattice weights
lattice-add-penalty: Add word insertion penalty to the lattice
lattice-best-path: Generate 1-best path through lattices; output as transcriptions and alignments
lattice-align-words: Convert lattices so that the arcs in the CompactLattice format correspond with words
lattice-align-phones:Convert lattices so that the arcs in the CompactLattice format correspond with phones