On the Recognition of Isolated Digits From a Large Telephone Customer Population

Abstract

A field study was initiated to learn about the effects of various telephone transmission and switching conditions on the algorithms currently used in the Bell Laboratories, Linear Predictive Coding (LPC)-based, isolated word recognizer. Digit recordings were obtained from customers over a variety of transmission facilities. During a 23-day recording period a total of 11,035 isolated digits were recorded. For each recording, statistics were recorded about the line condition, the background environment, and the customer's ability to speak his/her telephone number as a sequence of isolated digits. Also recorded was information about the ability of the automatic word end-point detector to find each spoken digit and to accurately determine the correct endpoints. The results of several recognition tests are presented — one using a previously defined set of laboratory-created digit reference templates, and several others using new sets of reference templates from a subset of the recorded digits. The performance of the recognizer is poor (average digit accuracy of 77.4 percent) using the laboratory template set, but improves substantially (average digit accuracy of 93.1 percent) for a template set created from the field recordings. The reasons for this improvement in digit recognition accuracy are presented, along with their implications to future work in isolated word recognition.