Isolated Word Recognition for Large Vocabularies

In the field of automatic speech recognition, the only type of system to date that has proven useful and practical is the isolated word recognizer. Isolated word recognizers have been in use commercially for a number of years,1"5 and have been extensively studied in several * Work performed on BLESP summer assignment at Bell Laboratories. 2989 major research laboratories throughout the world.6"13 For the most part, applications of isolated word recognizers have limited themselves to vocabulary sizes ranging from small (10 to 30 words) to moderate (30 to 200 words). There are several reasons why there are no commercially available systems that can recognize words from large vocabularies (greater than 200 words). These include: (i) The difficulty of training the system on large vocabularies (ii) The storage required for word templates for large vocabularies (Hi) The processing required to recognize words from large vocabularies (iv) The difficulty of accurately recognizing word vocabularies that would be useful in a variety of applications. The computational problems associated with reasons ii and iii above (i.e., large storage and large amounts of computation) are rapidly becoming less important as memory and processing costs decrease, and should continue to do so for the forseeable future. The problems in training are very real ones, and will be discussed further in this paper. The problems associated with choice of vocabulary words and accuracy of word recognition are the main topics of this paper.* Although the practicability of large vocabularies for isolated word recognition is doubtful, the experimental use of large vocabularies provides the opportunity to examine significant issues in automatic word recognition that cannot be examined with small vocabularies.