초록 ▼

A method is described for compressing the storage space required by HMM prototypes in an electronic memory. For this purpose prescribed HMM prototypes are mapped onto compressed HMM prototypes with the aid of a neural network (encoder). These can be stored with a smaller storage space than the uncom

A method is described for compressing the storage space required by HMM prototypes in an electronic memory. For this purpose prescribed HMM prototypes are mapped onto compressed HMM prototypes with the aid of a neural network (encoder). These can be stored with a smaller storage space than the uncompressed HMM prototypes. A second neural network (decoder) serves to reconstruct the HMM prototypes.

대표청구항 ▼

1. A method for compressing the storage space required by Hidden Markov Model (HMM) prototypes in an electronic memory, comprising:prescribing HMM prototypes; mapping the HMM prototypes onto compressed HMM prototypes using a first ancoder; and converting the compressed HMM prototypes into binary num

1. A method for compressing the storage space required by Hidden Markov Model (HMM) prototypes in an electronic memory, comprising:prescribing HMM prototypes; mapping the HMM prototypes onto compressed HMM prototypes using a first ancoder; and converting the compressed HMM prototypes into binary numbers using a second encoder, the second encoder being a bit encoder, the HMM prototypes being converted to binary numbers such that the binary numbers having as few bits as possible, and compressed HMM prototypes and the binary numbers being formed in such a way that a neural network can map the binary numbers onto reconstructed HMM prototypes, and the spacing between the HMM prototypes and the reconstructed HMM prototypes is minimized. 2. The method as claimed in claim 1, wherein the first encoder is a second neural network.3. The method as claimed in claim 1, wherein the reconstructed HMM prototypes with the higher ability to discriminate are weighted more heavily in order to determine the spacing between the HMM prototypes and the reconstructed HMM prototypes.4. The method as claimed in claim 2, wherein the reconstructed HMM prototypes with the higher ability to discriminate are weighted more heavily in order to determine the spacing between the HMM prototypes and the reconstructed HMM prototypes.5. The method as claimed in claim 1 whereinthe binary numbers have as few bits as possible such that the binary numbers occupy fewer bits than binary numbers obtained from a direct mapping. 6. A system for automatic speech recognition, comprising:a first encoder to map Hidden Markov Model (HMM) prototypes onto compressed HMM prototypes; a second encoder to convert the compressed HMM prototypes into binary numbers, the second encoder being a bit encoder; an electronic memory storage space to store the compressed HMM prototypes; a neural network to calculate speech patterns and to map the compressed HMM prototypes onto reconstructed HMM prototypes by mapping the binary numbers onto reconstructed HMM prototypes; an input to calculate a speech signal; a recognizer to recognize the input speech signal, the recognizer using the speech patterns calculated by the neural network ; and an output to produce a signal which corresponds to the input speech signal, wherein the compressed HMM prototypes and the neural network are chosen to minimize the spacing between the HMM prototypes and the reconstructed HMM prototypes, and the structure and parameters of the encoders and of the neural network are selected in such a way that the binary numbers have as few bits as possible. 7. The system as claimed in claim 6, wherein the memory stores compressed speech patterns calculated by the neural network.8. The system as claimed in claim 6, wherein the neural network comprises a single hidden layer.9. A computer readable storage medium storing a program to control a computer to perform a process for compressing the storage space required by Hidden Markov Model (HMM) prototypes, the process comprising:prescribing HMM prototypes (Xj) being prescribed; mapping the HMM prototypes onto compressed HMM prototypes using a first encoder; and converting the compressed HMM prototypes into binary numbers using a second encoder, the second encoder being a bit encoder, the HMM prototypes being converted to binary numbers such that the binary numbers having as few bits as possible, and the compressed HMM prototypes and the binary numbers being formed in such a way that the neural network can map the binary numbers onto reconstructed HMM prototypes, and the spacing between the HMM prototypes and the reconstructed HMM prototypes is minimized.