Lightly Supervised and Data-Driven Approaches to Mandarin Broadcast News Transcription

Computational Linguistics and Chinese Language Processing Vol. 10, No. 1, March 2005, pp. 1-18 1 © The Association for Computational Linguistics and Chinese Language Processing Lightly Supervised and Data-Driven Approaches to Mandarin Broadcast News Transcription Berlin Chen∗, Jen-Wei Kuo∗ and Wen-Hung Tsai∗ Abstract This article investigates the use of several lightly supervised and data-driven approaches to Mandarin broadcast news transcription. With the special structural properties of the Chinese language taken into consideration, a fast acoustic look-ahead technique for estimating the unexplored part of a speech utterance is integrated into lexical tree search to improve search efficiency. This technique is used in conjunction with the conventional language model look-ahead technique. Then, a verification-based method for automatic acoustic training data acquisition is proposed to make use of large amounts of untranscribed speech data. Finally, two alternative strategies for language model adaptation are studied with the goal of achieving accurate language model estimation. With the above approaches, the overall system was found in experiments to yield an 11.88% character error rate when applied to Mandarin broadcast news collected in Taiwan. Keywords: acoustic look-ahead, lightly supervised acoustic model training, language model adaptation, Mandarin broadcast news 1. Introduction With the continuing growth of the amount of multimedia information accessible over the Internet, large volumes of real-world speech information, such as that in broadcast radio and television programs, digital libraries, and so on, are now being accumulated and made available to the public. Substantial efforts and very encouraging results for broadcast news transcription, retrieval, and summarization have been reported [Woodland 2002; Gauvain et al. 2002; Beyerlein et al. 2002; Chen et al. 2002; Chang et al. 2002; Meng et al. 2004; Furui et al. 2004]. However, in order to obtain better recognition performance, most of the transcription systems require not only large amounts of manually transcribed speech materials for acoustic training in the data preparation phase, but also much time and memory in the recognition ∗ Graduate Institute of Computer Science and Information Engineering, National Taiwan Normal University, Taipei, Taiwan, Republic of China E-mail: {berlin, rogerkuo, louis}@csie.ntnu.edu.tw 2 Berlin Chen et al. phase. Moreover, because the subject domains and lexical regularities of the linguistic contents of news articles are very diverse and often change with time, it is extremely difficult to build well-estimated language models for speech recognition. Hence, in the recent past, several attempts have been made to investigate the possibility of achieving automatic acquisition of speech or language training data for system refinement or for rapid prototyping of a new recognition system to new domains, and very encouraging results have been obtained [Kemp and Waibel 1999; Wessel and Ney 2001; Macherey and Ney 2002; Bacchiani 2003]. On the other hand, quite a few studies have also explored ways to improve recognition efficiency, and many good approaches have been proposed [Schuster 2000; Aubert 2002; Evermann and Woodland 2003]. In this paper, several lightly supervised and data-driven approaches to Mandarin broadcast news transcription are presented. First, considering the special structural properties of the Chinese language, a …

Content truncated to 500 words