Large vocabulary Mandarin Final recognition based on Two-Level Time-Delay Neural Networks (TLTDNN)

Abstract

A Two-Level Time-Delay Neural Network (TLTDNN) technique has been developed to recognize all Mandarin Finals of the entire Chinese syllables. The first level discriminates the vowel-group based on (a, e, i, o, u, v) and the nasal-group based on nasal ending (-n, -ng, -others). The nasal-group discriminator is used to further split the large /a/ subgroup produced by the vowel-group discriminator. The two groupings in the first level produce 8 subgroups in the second level. Further discrimination in the second level enables the identification of all 35 Mandarin Finals. The technique was thoroughly tested using 8 sets of 1265 isolated Hanyu Pinyin syllables, with 6 sets used for training and 2 sets used for testing. The overall result shows that a high recognition rate of 99.4% on the training datasets and 95.6% on the test datasets, is achievable. The top 4 recognition rate attained on the test datasets is as high as 99.1%. © 1997 Elsevier Science B.V.