First International Conference on Spoken Language Processing (ICSLP 90)

Kobe, Japan
November 18-22, 1990

Speaker Independent Speech Recognition Based on Neural Networks of Each Category With Embedded Eigenvectors

Yasuyuki Masai, Hiroshi Matsu'ura, Tsuneo Nitta

Information & Communication Systems Laboratory, TOSHIBA Corporation, Kawasaki, Japan

This paper describes a speaker independent word recognition algorithm which is based on four layered neural networks with embedded eigenvectors. Eigenvectors in the Subspace Method (SM) are used as weights. In the SM, the accumulation of projection component values from an input pattern is used as a measure of similarity. In contrast to this, our proposed method utilizes each projection component value to achieve performance better than that of the SM. We propose the Subspace Training (SST) algorithm with the SM and the Decision Controlled Back Propagation Training (DCBPT) algorithm to reduce training times. Training and recognition experiments were performed using a 26 word vocabulary consisting of train station names. The error rate of the SM was 1.3%. The error rate was reduced to 0.7% using the neural networks combined with the SM.

Full Paper

Bibliographic reference.  Masai, Yasuyuki / Matsu'ura, Hiroshi / Nitta, Tsuneo (1990): "Speaker independent speech recognition based on neural networks of each category with embedded eigenvectors", In ICSLP-1990, 681-684.