ニュース音声認識のための時期依存言語モデル  [in Japanese] Time Dependent Language Model for Broadcast News Transcription  [in Japanese]

Access this Article

Search this Article

Author(s)

Abstract

本論文では ニュース番組の音声認識の性能を向上させるため 最新ニュースの内容に依存した言語モデル(時期依存言語モデル)を提案する. まず ニュース音声認識のための言語モデルについて 学習期間 学習時期を調べた. その結果 学習データは長期間であるほど また 評価時期に近いほど有効であることが分かった. そこで 長期間の大量のニュース原稿に対して 最新の小量のニュース原稿をMAP推定に基づいて重み付け混合し 時期依存言語モデルを作成した. 長期間ニュース原稿に対する最新ニュース原稿の重みは EMアルゴリズムにより推定される. 提案手法の特徴は n-gramを最新ニュースに適応化するだけではなく 語彙を自動的に更新することである. 認識実験の結果 時期依存言語モデルはテストセット297文中 最新ニュースに関連する140文について 単語正解精度を2.0%向上させ 未知語の延べ数を29.4%削減した. 一方 関連のないニュースについて単語正解精度は低下しなかった.In this paper we propose a new linguistic technique to make an adapted language model (Time Dependent Language Model, TDLM) which gives an improvement in broadcast news transcription. Examining a good condition of a language model for broadcast news transcription, we found out that long-term training scripts and the latest news scripts were effective. In our proposed method, the TDLM is trained from long-term news scripts and short-term latest news scripts based on MAP estimation. The mixture weight for short-term news scripts are derived from the EM algorithm. The method features not only an adaptation of n-grams for the latest news scripts but updating of TDLM's vocabulary automatically. we had a transcribing experiment for 297 sentences in NHK's broadcast news. The TDLM achieved 2.0% improvement in word accuracy over the baseline (no adapted) language model for 140 sentences which are associated with the latest news content and reduced 29.4% of unknown words in these topic related sentences. On the other hand, it didn't lower the word accuracy for the rest sentences.

In this paper we propose a new linguistic technique to make an adapted language model (Time Dependent Language Model, TDLM) which gives an improvement in broadcast news transcription. Examining a good condition of a language model for broadcast news transcription, we found out that long-term training scripts and the latest news scripts were effective. In our proposed method, the TDLM is trained from long-term news scripts and short-term latest news scripts based on MAP estimation. The mixture weight for short-term news scripts are derived from the EM algorithm. The method features not only an adaptation of n-grams for the latest news scripts but updating of TDLM's vocabulary automatically. we had a transcribing experiment for 297 sentences in NHK's broadcast news. The TDLM achieved 2.0% improvement in word accuracy over the baseline (no adapted) language model for 140 sentences which are associated with the latest news content and reduced 29.4% of unknown words in these topic related sentences. On the other hand, it didn't lower the word accuracy for the rest sentences.

Journal

  • Transactions of Information Processing Society of Japan

    Transactions of Information Processing Society of Japan 40(4), 1421-1429, 1999-04-15

    Information Processing Society of Japan (IPSJ)

References:  14

Cited by:  51

Codes

  • NII Article ID (NAID)
    110002724796
  • NII NACSIS-CAT ID (NCID)
    AN00116647
  • Text Lang
    JPN
  • Article Type
    Journal Article
  • ISSN
    1882-7764
  • NDL Article ID
    4692806
  • 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.
    Z14-741
  • Data Source
    CJP  CJPref  NDL  NII-ELS  IPSJ 
Page Top