Parameter tying for flexible speech recognition

The paper presents two parameter tying techniques which enable a trade-off between computational cost and recognition performances of a speaker independent flexible speech recognition system working over the telephone network. Parameter tying is conducted at phonetic and acoustic levels. At the phonetic level, allophone and triphone based phonetic modeling are used simultaneously to achieve the best trade-off between computational cost and recognition performances. This decreases error rate with a controlled computational cost as compared to an allophone modeling. At the acoustic level, the tying is performed by clustering the Gaussian densities of mixture distributions. After clustering, a particular density may be use by several distribution. This allows the total number of Gaussian densities to be divided by two while improving the recognition performances