Arithmetic logic unit architecture
View Patent ↗Embodiments of the present invention include an apparatus, method, and system for acoustic modeling. In an embodiment, an arithmetic logic unit for computing a one-dimensional score between a feature vector and a Gaussian probability distribution vector is provided. The arithmetic logic unit includes a computational logic unit configured to compute a first value based on a mean value and a variance value associated with a dimension of the Gaussian probability distribution vector and a dimension of a feature vector, a look up table module configured to output a second value based on the variance value, and a combination module configured to combine the first value and the second value to generate the one-dimensional score.
1. An arithmetic logic unit for computing a one-dimensional score between a feature vector and a Gaussian probability distribution vector, comprising:
a computational logic unit configured to compute a first value based on a mean value and a variance value associated with a dimension of the Gaussian probability distribution vector and a dimension of a feature vector;
a look up table module configured to output a second value based on the variance value; and
a combination module configured to combine the first value and the second value to generate the one-dimensional score.
2. The arithmetic logic unit of claim 1 , wherein the first value and the second value are represented using a same number of bits.
3. The arithmetic logic unit of claim 1 , wherein the computational logic unit comprises a subtraction module configured to compute the difference between the one-dimension value of the feature vector and the mean value and wherein the difference is positive.
4. The arithmetic logic unit of claim 3 , wherein the computational logic unit further comprises a look up table configured to output a third value based on the variance value.
5. The arithmetic logic unit of claim 4 , wherein the computational logic unit is configured to compute the first value as a product of a square of the difference and the third value.
6. The arithmetic logic unit of claim 1 , wherein the computational logic unit comprises an exception handling module configured to output the first value for specific values of the mean value, the variance value, and the dimension of the feature vector.
7. The arithmetic logic unit of claim 1 , wherein the computational logic unit comprises a feature vector transformation module configured to scale the dimension of the feature vector by a scaling factor.
8. The arithmetic logic unit of claim 1 , wherein the combination module comprises:
an adder configured to compute a sum of the first and second values; and
a shift module configured to shift the sum such that 0 bits are present right of a decimal point of the sum.
9. The arithmetic logic unit of claim 1 , wherein the lookup table module comprises a look up table configured to output a value lut ij given by
lut ij =(ln(var ij )− C )× M 1 ,
where:
var ij , is the variance value of the i th dimension of the j th Gaussian probability distribution vector; and
M 1 is a scaling factor; and
C is a constant.
10. The arithmetic logic unit of claim 1 , wherein the first value is given by:
(
(
x
i
-
μ
ij
)
×
(
1
var
ij
)
)
2
×
M
2
,
where
x i is the value of the i th feature vector dimension,
μ ij is the mean value of the i th dimension of the j th Gaussian probability distribution vector;
var ij , is the variance value of the i th dimension of the j th Gaussian probability distribution vector; and
M 2 is a weighting factor.
11. A method of computing a one-dimensional distance score between a feature vector and a Gaussian probability distribution vector, comprising:
computing a first value based on a mean value and a variance value associated with a dimension of the Gaussian probability distribution vector and a dimension of a feature vector;
retrieving a second value based on the variance value from a look up table module; and
combining the first value and the second value to generate the one-dimensional score; and
wherein at least one of the computing, retrieving, and combining are performed by one or more computers.
12. The method of claim 11 , wherein the computing comprises:
computing a difference between the dimension of the feature vector and the mean value.
13. The method of claim 12 , wherein the computing comprises retrieving a third value from a second look up table based on the variance value.
14. The method of claim 13 , wherein the computing comprises computing the first value as a product of a square of the difference and the third value.
15. The method of claim 11 , further comprising, before the computing, transforming the value of the feature vector module for that dimension using a scaling factor.
16. The method of claim 11 , wherein the lookup table module comprises a look up table configured to output a value lut ij given by
lut ij =(ln(var ij )− C )× M 1 ,
where:
var ij , is the variance value of the i th dimension of the j th Gaussian probability distribution vector;
M 1 is a scaling factor; and
C is a constant.
17. The method of claim 11 , wherein the first value is given by:
(
(
x
i
-
μ
ij
)
×
(
1
var
ij
)
)
2
×
M
2
,
where:
x i is the value of the i th feature vector dimension,
μ ij is the mean value of the i th dimension of the j th Gaussian probability distribution vector;
var ij , is the variance value of the i th dimension of the j th Gaussian probability distribution vector; and
M 2 is a weighting factor.