Unsupervised Language Acquisition Integrating Bayesian Double Articulation Analyzer and Deep Learning

2018 Fiscal Year Final Research Report

• Project/Area Number: 15H05319
• Research Category: Grant-in-Aid for Young Scientists (A)
• Allocation Type: Single-year Grants
• Research Field: Soft computing
• Research Institution: Ritsumeikan University
• Principal Investigator: Taniguchi Tadahiro 立命館大学, 情報理工学部, 教授 (80512251)
• Project Period (FY): 2015-04-01 – 2019-03-31
• Keywords: 記号創発 / 言語獲得 / ノンパラメトリックベイズ / 深層学習 / 音声認識

Outline of Final Research Achievements

In this study, we extended the Bayesian double articulation analyzer aiming at achieving automatic phoneme and word discovery from real human speech. There were three main achievements. (1) We have realized the improvement of the accuracy of unsupervised phoneme and word acquisition from actual speech data by utilizing deep learning. (2) We developed an accelerated Bayesian double articulation analyzer that is 100 times faster than the previous algorithm. (3) We developed an unsupervised language acquisition method from multiple speakers by developing unsupervised speaker adaptation method using deep learning.

Free Research Field

創発システム

Academic Significance and Societal Importance of the Research Achievements

ロボットや人工知能が音声言語を認識するためには従来，大量の音声データとその書き起こし文を事前に人手で準備し，機械学習によって音声認識装置を訓練する必要があった．しかし，これは人間の幼児の言語獲得過程と一致しない．人間の幼児は書き言葉を学ぶ前に，音声データから音素や単語を発見していく．この過程を実現するために，本研究では研究代表者が構築してきたベイズ二重分節解析器に深層学習を組み合わせて，高い精度での音素・単語獲得，複数話者からの学習，また，そのアルゴリズムの高速化を実現した．