7th International Conference on Spoken Language Processing
September 16-20, 2002
In the acoustic modeling for large vocabulary speech recognition, context-dependent (CD) modeling is essential for realizing both improved recognition performance and rapid search. However, sparse data problem caused by huge number of CD models usually leads the estimated models unreliable. To cope with that, two major contextclustering methods, data-driven and rule-based, have been investigated vigorously. In this paper, we briefly review the two methods and develop a new clustering method based on ID3 decision tree learning algorithm that effectively captures the CD modeling. The proposed scheme essentially constructs a decision rule of preclustered triphones using ID3 algorithm. In particular, the datadriven method is used as a clustering algorithm while its result is used as the learning target of ID3 algorithm. The proposed scheme is shown effective over the database of low unknown-context ratio in terms of recognition performance. For speaker-independent, taskindependent continuous speech recognition task, the proposed method reduced percent accuracy WER by 1.16% comparing to that of the existing rule-based method alone.
Bibliographic reference. Park, Junho / Ko, Hanseok (2002): "Construction of decision tree from data driven clustering", In ICSLP-2002, 2657-2660.