大語彙連続音声認識エンジンJuliusにおける A * 探索法の改善  [in Japanese] Improvements of A * - based search algorithm in LVCSR engine Julius  [in Japanese]

Access this Article

Search this Article

Author(s)

    • 李晃伸 LEE Akinobu
    • 京都大学大学院 情報学研究科 知能情報学専攻 Graduate School of Informatics, Kyoto University

Abstract

大語彙連続音声認識エンジンJuliusにおける解探索アルゴリズムの種々の改善手法を提案し,その評価を行う.ヒューリスティックの非適格性から生じる探索誤りや探索失敗を解消するために,第2パスで探索が幅優先に陥るのを防ぐenveloped best-first探索を提案するとともに,第1パスで単語間triphoneを近似計算することで探索の高精度化を図る.高速化の面からは,第2パスの音響尤度計算におけるスコアでのビーム設定と,第1パスでの1-gram確率に基づくfactoringを導入する.JNASの20 000語タスクでの評価実験の結果探索誤りの多くが解消され,またその精度を落とさずに計算量を削減することができた.最終的に,実時間の12.9倍で94.9%,monophoneではほぼ実時間で84.2%の単語認識精度を得ることができた.The recent improvements in our LVCSR engine "Julius" are shown. To ease search errors and search failure caused by dis-optimality of heuristics, an enveloped best-first search algorithm and an approximation of the inter-word context dependency on the 1st pass are proposed. More, score-based envelope beam for aoustic scanning on the 2nd pass and 1-gram factoring are introduced to decrease computational costs. Experiments on 20,000-word JNAS task show that most of the search errors are dissolved and the costs can be efficiently cut with little accuracy loss. The system achieved word accuracy of 94.9% in a real-time factor of 12.9, and using monophone model, 84.2% in nearly real-time.

The recent improvements in our LVCSR engine "Julius" are shown. To ease search errors and search failure caused by dis-optimality of heuristics, an enveloped best-first search algorithm and an approximation of the inter-word context dependency on the 1st pass are proposed. More, score-based envelope beam for aoustic scanning on the 2nd pass and 1-gram factoring are introduced to decrease computational costs. Experiments on 20,000-word JNAS task show that most of the search errors are dissolved and the costs can be efficiently cut with little accuracy loss. The system achieved word accuracy of 94.9% in a real-time factor of 12.9, and using monophone model, 84.2% in nearly real-time.

Journal

  • IPSJ SIG Notes

    IPSJ SIG Notes 1999(64(1999-SLP-027)), 33-40, 1999-07-23

    Information Processing Society of Japan (IPSJ)

References:  9

Cited by:  20

Codes

  • NII Article ID (NAID)
    110002917103
  • NII NACSIS-CAT ID (NCID)
    AN10442647
  • Text Lang
    JPN
  • Article Type
    Journal Article
  • ISSN
    09196072
  • NDL Article ID
    5338297
  • NDL Source Classification
    ZM13(科学技術--科学技術一般--データ処理・計算機)
  • NDL Call No.
    Z14-1121
  • Data Source
    CJP  CJPref  NDL  NII-ELS  IPSJ 
Page Top