Design of speech recognition system: problems and solutions

✍ Scribed by Takao Watanabe

• Publisher: John Wiley and Sons
• Year: 1999
• Tongue: English
• Weight: 551 KB
• Volume: 30
• Category: Article
• ISSN: 0882-1666
• DOI: 10.1002/(sici)1520-684x(19991115)30:12<95::aid-scj10>3.0.co;2-8

✦ Synopsis

This paper discusses a number of problems that are to be solved for the purpose of practicable implementation of speech recognition systems, and a method of speech recognition is proposed using the authors concept of demisyllable units. In the proposed implementation of demisyllable-based speaker-independent continuous speech recognition, such improved robustness features as speaker adaptation using spectral interpolation mapping, environment adaptation using the REALISE method, and a method of unknown input rejection based on likelihood correction using syllable recognition were provided. Further, acceleration of likelihood calculation using a tree-structured probability distribution was proposed, as well as cost reduction by acceleration of continuous speech recognition using the bundle search. Software development was also described. The proposed method is expected to contribute to the practicability of speech recognition, although many problems, particularly those related to system robustness, still remain to be solved.