Matrix quantization with vector quantization error compensation and neural network postprocessing for robust speech recognition
First Claim
1. A speech recognition system comprising:
- a vector quantizer to receive first parameters of an input signal and to generate a first quantization observation sequence;
a first speech classifier to receive the first quantization observation sequence from the vector quantizer and to generate first respective speech classification output data;
a matrix quantizer to receive second parameters of the input signal, and to generate a second quantization observation sequence;
a second speech classifier to receive the second quantization observation sequence from the matrix quantizer and to generate second respective speech classification output data;
a mixer to combine corresponding first and second respective speech classification data to generate third respective speech classification data and to generate output data from the first, second, and third speech classification data; and
a neural network to receive output data from the mixer and to determine fourth respective speech classification output data.
11 Assignments
0 Petitions
Accused Products
Abstract
A speech recognition system utilizes both matrix and vector quantizers as front ends to a second stage speech classifier such as hidden Markov models (HMMs) and utilizes neural network postprocessing to, for example, improve speech recognition performance. Matrix quantization exploits the “evolution” of the speech short-term spectral envelopes as well as frequency domain information, and vector quantization (VQ) primarily operates on frequency domain information. Time domain information may be substantially limited which may introduce error into the matrix quantization, and the VQ may provide error compensation. The matrix and vector quantizers may split spectral subbands to target selected frequencies for enhanced processing and may use fuzzy associations to develop fuzzy observation sequence data. A mixer provides a variety of input data to the neural network for classification determination. The neural network'"'"'s ability to analyze the input data generally enhances recognition accuracy. Fuzzy operators may be utilized to reduce quantization error. Multiple codebooks may also be combined to form single respective codebooks for split matrix and split vector quantization to reduce processing resources demand.
87 Citations
44 Claims
-
1. A speech recognition system comprising:
-
a vector quantizer to receive first parameters of an input signal and to generate a first quantization observation sequence;
a first speech classifier to receive the first quantization observation sequence from the vector quantizer and to generate first respective speech classification output data;
a matrix quantizer to receive second parameters of the input signal, and to generate a second quantization observation sequence;
a second speech classifier to receive the second quantization observation sequence from the matrix quantizer and to generate second respective speech classification output data;
a mixer to combine corresponding first and second respective speech classification data to generate third respective speech classification data and to generate output data from the first, second, and third speech classification data; and
a neural network to receive output data from the mixer and to determine fourth respective speech classification output data. - View Dependent Claims (2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16)
-
-
17. A speech recognition system comprising:
-
a vector quantizer to receive first parameters of an input signal and to generate a first quantization observation sequence, wherein the first parameters are grouped into S1 partition(s);
a split matrix quantizer to receive second parameters of the input signal and to generate a second quantization observation sequence, wherein the second parameters are grouped into S2 partition(s);
a first speech classifier to receive the first quantization observation sequence from the vector quantizer and generate first respective speech classification output data;
a second speech classifier to receive the second quantization observation sequence from the split matrix quantizer and generate second respective speech classification output data;
a mixer to combine corresponding first and second respective speech classification data to generate third respective speech classification data and to provide output data based on the first, second, and third classification data; and
a neural network to receive the mixer output data and to generate fourth respective speech classification data based on the mixer output data. - View Dependent Claims (18, 19)
-
-
20. An apparatus comprising:
-
a first speech classifier to operate on S1 group(s) of first parameters of an input signal and to provide first output data relating the input signal to first reference data, wherein the first input signal parameters include frequency and time domain parameters, wherein S1 is a positive integer;
a second speech classifier to operate on S2 group(s) of second parameters of the input signal and to provide second output data relating the second input signal to second reference data, wherein the second parameters of the input signal include the frequency domain parameters, wherein S2 is a positive integer;
mixer to combine the first output data and the second output data into third output data so that the second output data compensates for errors in the first output data; and
a neural network to receive selected output data from the mixer and to generateoutput data to classify the input signal. - View Dependent Claims (21, 22, 23, 24, 25, 26, 27, 28, 29)
wherein at least one of the S2 partitions of the second parameters of the input signal are corrupted by noise and the respective distance measure to relate the respective noise corrupted second parameters to partitioned second reference data has noise rejection features.
-
-
24. The apparatus as in claim 20 wherein S1 is greater than one and S2 is greater than one.
-
25. The apparatus as in claim 20 wherein the first speech classifier includes a fuzzy split matrix quantizer, and the second speech classifier includes a fuzzy split vector quantizer.
-
26. The apparatus as in claim 25 wherein the first speech classifier further includes a first set of hidden Markov models, and the second speech classifier further includes a second set of hidden Markov models.
-
27. The apparatus as in claim 20 wherein the second speech classifier is capable of operating on frequency domain parameters of the input signal.
-
28. The apparatus as in claim 20 wherein the frequency domain parameters are P order line spectral pair frequencies, wherein P is an integer.
-
29. The apparatus as in claim 20 wherein the first and second parameters of the input signal further include input signal energy related parameters.
-
30. A method comprising the steps of:
-
processing first parameters of the input signal to relate the first parameters to first reference data wherein the first parameters include frequency and time domain information;
generating first output data relating the first parameters to reference data;
processing second parameters of the input signal to relate the second parameters to second reference data wherein the second parameters include frequency domain information;
generating second output data relating the second parameters to the second reference data;
combining the first output data and second output data into third output data to compensate for errors in the first output data; and
providing the first, second, and third output data to a neural network to classify the input signal. - View Dependent Claims (31, 32, 33, 34, 35, 36, 37, 38, 39)
partitioning the first parameters of the input signal into S1 groups; and
partitioning the second parameters of the input signal into S2 groups.
-
-
32. The method as in claim 31 wherein the step of partitioning first parameters of an input signal into S1 groups comprises the step of:
partitioning the first parameters of the input signal to group at least one subset of the first parameters which are generally corrupted by localized noise.
-
33. The method as in claim 32 wherein the step of partitioning first parameters of an input signal into S1 groups comprises the step of:
partitioning the first parameters of the input signal to group at least one subset of the first parameters which are generally corrupted by localized noise.
-
34. The method as in claim 30 wherein the first parameters and first reference data include respective corresponding line spectral pair frequencies, the second parameters and second reference data include respective corresponding line spectral pair frequencies, and the subset of the first parameters which are generally corrupted by localized noise are the mththrough nthline spectral frequencies, the step of processing the first parameters further comprising the step of:
-
matrix quantizing the mththrough nthline spectral frequencies of the first parameters using a distance measure proportional to (i) a difference between the ithinput signal line spectral pair frequencies and the ithorder first reference data line spectral pair frequencies and (ii) a weighting of the difference by an ithfrequency weighting factor, wherein m is less than or equal to i, and n is greater than or equal to i; and
the step of processing the second parameters further comprising the step of;
vector quantizing the mththrough nthline spectral frequencies of the second parameters using a distance measure proportional to (I) a difference between the ithinput signal line spectral pair frequencies and the ithorder second reference data line spectral pair frequencies and (ii) a weighting of the difference by an ithfrequency weighting factor, wherein m is less than or equal to i.
-
-
35. The method as in claim 30 wherein the step of processing the first parameters of the input signal comprises the step of:
-
matrix quantizing each of the partitioned first parameters of the input signal; and
the step of processing second parameters of the input signal comprises the step of;
vector quantizing each of the second parameters of the input signal.
-
-
36. The method as in claim 35 wherein the step of matrix quantizing further comprises the step of:
-
fuzzy matrix quantizing each of the first parameters of the input signal; and
wherein the step of vector quantizing further comprises the step of;
fuzzy vector quantizing each of the second parameters of the input signal.
-
-
37. The method as in claim 36 wherein the step of fuzzy matrix quantizing further comprises the step of:
-
fuzzy matrix quantizing each of the first parameters of the input signal using a first codebook; and
wherein the step of fuzzy vector quantizing further comprises the step of;
fuzzy vector quantizing each of the second parameters of the input signal using a second single codebook.
-
-
38. The method as in claim 35 wherein the step of processing the first parameters of the input signal further comprises the step of:
-
determining first respective input signal recognition probabilities from a plurality of first hidden Markov models; and
wherein the step of processing the second parameters of the input signal further comprises the step of;
determining second respective input signal recognition probabilities from a plurality of second hidden Markov models.
-
-
39. The method as in claim 30 wherein the step of combining comprises the step of:
-
weighting the second output data; and
adding the weighted second output data to the first output data.
-
-
40. A method of recognizing speech comprising the steps of:
-
receiving an input signal;
determining parameters of the input signal;
vector quantizing the parameters of the input signal to obtain first quantization output data;
classifying the first quantization output data;
matrix quantizing the parameters of the input signal to obtain second quantization output data;
classifying the second quantization output data;
combining the first and second quantization output data to generate third output data; and
generating an identification of the input signal with a neural network based upon the classification of the first and second quantization output data and the third output data. - View Dependent Claims (41, 42, 43)
weighting the classification of the first quantization output data; and
adding the weighted classification of the first quantization output data and the classification of the second quantization output data.
-
-
42. The method as in claim 40 wherein the step of determining parameters of the input signal comprises the step of:
determining P order line spectral pairs for each of TO frames of the input signal.
-
43. The method as in claim 40 wherein the step of vector quantizing further comprises the step of:
-
fuzzy split vector quantizing the parameters of the input signal, wherein the first quantization output data is fuzzy data; and
wherein the step of matrix quantizing further comprises the step of;
fuzzy split matrix quantizing the parameters of the input signal, wherein the second quantization output data is fuzzy data.
-
-
44. A method of recognizing speech comprising the steps of:
-
receiving an input signal;
determining D order line spectral pairs for TO frames of the input signal, wherein D and TO are integers;
determining parameters related to the energy of the input signal, wherein the parameters related to the energy of the input signal include the input signal energy and a first derivative of the input signal energy;
vector quantizing the D order line spectral pairs for each of the TO frames and the parameters related to the input signal energy;
classifying the input signal using the vector quantization of the D order line spectral pairs;
matrix quantizing the D order line spectral pairs and the parameters related to the input signal energy for T matrices of frames of the input signal, wherein T is defined as int(TO/N), and N is the number for input signal frames represented in each of the T matrices;
classifying the input signal using the matrix quantization of the D order line spectral pairs and parameters related to the input signal energy;
combining the classifications of the input signal and providing the individual classifications of the input signal and the combined classification of the input signal to a neural network.
-
Specification