Sub-partitioned vector quantization of probability density functions
First Claim
1. A method for creating a subpartitioned vector quantized memory for the storage of hidden Markov model (HMM) log-probability density functions (log-pdfs) corresponding to a phoneme model having at least one code-book and one state, comprising the following steps:
- a) organizing the HMM log-pdfs of each code-book by column and grouped by state so that corresponding log-pdf values of each of the HMM log-pdfs form a set of log-pdf value columns;
b) subpartitioning the log-pdf value columns into an integer number of equal length packets each packet identified by an associated packet index;
c) vector quantizing the subpartitioned packets, creating a set of subpartitioned vector quantization (SVQ) encoding vectors and associated SVQ encoding vector indices;
d) constructing an address translation table that is addressable by the packet indices, listing the SVQ encoding vector indices associated with each packet index, for generating, at output, an encoding index corresponding to the packet index used to address the address translation table; and
e) constructing a SVQ vector table for storing the set of SVQ encoding vectors in accordance with the associated SVQ encoding vector indices.
3 Assignments
0 Petitions
Accused Products
Abstract
A speech recognition memory compression method and apparatus subpartitions probability density function (pdf) space along the hidden Markov model (HMM) index into packets of typically 4 to 8 log-pdf values. Vector quantization techniques are applied using a logarithmic distance metric and a probability weighted logarithmic probability space for the splitting of clusters. Experimental results indicate a significant reduction in memory can be obtained with little increase in overall speech recognition error.
83 Citations
25 Claims
-
1. A method for creating a subpartitioned vector quantized memory for the storage of hidden Markov model (HMM) log-probability density functions (log-pdfs) corresponding to a phoneme model having at least one code-book and one state, comprising the following steps:
-
a) organizing the HMM log-pdfs of each code-book by column and grouped by state so that corresponding log-pdf values of each of the HMM log-pdfs form a set of log-pdf value columns; b) subpartitioning the log-pdf value columns into an integer number of equal length packets each packet identified by an associated packet index; c) vector quantizing the subpartitioned packets, creating a set of subpartitioned vector quantization (SVQ) encoding vectors and associated SVQ encoding vector indices; d) constructing an address translation table that is addressable by the packet indices, listing the SVQ encoding vector indices associated with each packet index, for generating, at output, an encoding index corresponding to the packet index used to address the address translation table; and e) constructing a SVQ vector table for storing the set of SVQ encoding vectors in accordance with the associated SVQ encoding vector indices. - View Dependent Claims (2, 3, 4, 5, 6, 7, 8, 9)
-
-
10. A subpartitioned vector quantization memory compression storage and retrieval system for HMM log-pdfs, addressable by a read address that includes a packet index corresponding to a packet location in uncompressed memory, comprising:
-
a) an address translation table memory for storing and outputting SVQ encoding vector indices at an output port, each addressable through a read address port by a corresponding packet index; and b) a vector table memory for storing a set of SVQ encoding vectors, addressable by a corresponding encoding vector index, with a read address port connected to the output of the address translation table memory, and an output port for providing the encoded vector corresponding to a packet specified by the packet index. - View Dependent Claims (11, 12)
-
-
13. A method for creating a subpartitioned vector quantized (SVQ) memory for compressing capacity memory for the storage of a set of discrete log-probability density functions (log-pdfs) each with an equal number of prescribed elements, comprising the following steps:
-
a) arranging the set of log-pdfs as a matrix wherein each log-pdf of the set is contained in a row and the elements of each row are elements of distinct columns; b) subpartitioning the log-pdf value columns into an integer number of equal length packets each packet identified by an associated packet index; c) vector quantizing the subpartitioned packets, creating a set of subpartitioned vector quantization (SVQ) encoding vectors and associated SVQ encoding vector indices; d) constructing an address translation table that is addressable by the packet indices, listing the SVQ encoding vector indices associated with each packet index, for generating, at output, an encoding index corresponding to the packet index used to address the address translation table; and e) constructing a SVQ vector table for storing the set of SVQ encoding vectors in accordance with the associated SVQ encoding vector indices. - View Dependent Claims (14, 15, 16, 17, 18, 19, 20, 21, 22)
-
-
23. A speech recognition system comprising:
-
a) a speech transducer for generating an electrical signal representative of the input acoustical speech signal; b) an analog-to-digital converter for scalar quantization of the electrical signal at its output, having an input port connected to the output of the speech transducer; c) a speech signal feature extraction processor connected to the output of the analog-to-digital converter for extracting a speech feature vector; d) a vector quantizer having an input connected to the output of the feature extraction processor for producing a vector quantized speech feature vector at an output; e) a phoneme probability processor connected to the output of the vector quantifier for operating on the vector quantized speech feature vector and computing a set of probabilities that a given hidden Markov model produced the speech feature vector based on a prescribed set of hidden Markov models; f) a hidden Markov model memory for storing a prescribed set of hidden Markov models which is implemented as a sub-partitioned vector quantization storage and retrieval system, addressable by the phoneme probability processor and for producing at an output phone probabilities, the output connected to the phoneme probability processor; and g) a search engine for searching for a candidate sentence, given the phone probabilities, and for outputting a word sequence identifier for the most probable word sequence of the candidate sentence. - View Dependent Claims (24, 25)
-
Specification