Some experiments of a speaker‐independent isolated word recognition using the Selective Weighted Matching

Abstract

At the last meeting in Chicago, we reported the speaker‐independent isolated word recognition system using the Selective Weighted Matching. This paper refers to the results of some experiments to investigate the relation between recognition accuracy and following items: (1) the number of training talkers, (2) the number of reference patterns, and (3) the recognition accuracy for 100 test talkers. In these experiments, we used a vocabulary (digits and two command words in Japanese) uttered over a telephone line by 500 talkers. Concerning (1), we made recognition tests for 100 test talkers within 96 reference patterns. The number of training talkers was varied from 100 to 400. In the case of 100 talkers, the recognition accuracy was 96.96% for the training talkers and 93.37% for the test talkers. In the case of 400 talkers, the results were 95.37% and 94.17%, respectively. By increasing the training talkers, the recognition accuracy was inclined to decrease for the training talkers and to increase for the test talkers, and the difference between them was diminished. From the results of these experiments, we decided the practical conditions for our recognition system.