摘要
To efficiently decrease the size of parameters and improve the robustness of parameters training, a fuzzy clustering based phonetic tied-mixture model, FPTM, is presented. The Gaussian codebook of FPTM is synthesized from Gaussian components belonging to the same root node in phonetic decision tree. Fuzzy clustering method is further used for FPTM covariance sharing. Experimental results show that compared with the conventional PTM with approximately the same parameters size, FPTM decrease the size of Gaussian weights by 77.59% and increases word accuracy by 7.92%, which proves Gaussian fuzzy clustering is efficient. Compared with FPTM, covariance-shared FPTM decreases word error rate by 1.14% , which proves the combined fuzzy clustering for both Gaussian and covariance is superior to Gaussian fuzzy clustering alone.
To efficiently decrease the size of parameters and improve the robustness of parameters training, a fuzzy clustering based phonetic tied-mixture model, FPTM, is presented. The Gaussian codebook of FPTM is synthesized from Gaussian components belonging to the same root node in phonetic decision tree. Fuzzy clustering method is further used for FPTM covariance sharing. Experimental results show that compared with the conventional PTM with approximately the same parameters size, FPTM decrease the size of Gaussian weights by 77.59% and increases word accuracy by 7.92%, which proves Gaussian fuzzy clustering is efficient. Compared with FPTM, covariance-shared FPTM decreases word error rate by 1.14% , which proves the combined fuzzy clustering for both Gaussian and covariance is superior to Gaussian fuzzy clustering alone.
基金
The work was supported by the Science and Technology Committee of Shanghai (01JC14033).
