Method and apparatus for processing audio and speech signals
A method and device for processing signals representing speech or audio via a plurality of filters that approximate behaviors of the basilar membrane of human cochlea. Each of the plurality of filters is formed from a mother filter via the dilation and a shift in time and has the similar impulse response of the basilar membrane to the frequency band for which the filter represents. Any process can be conducted and any feature can be extracted in the domain of the filters' outputs for applications, such as noise reduction, speech synthesis, coding, and speech and speaker recognition. Processed signals can be synthesized back to the time domain via an inverse cochlear transform.
1. A method for processing signals representing speech or audio via a plurality of filters that approximate behaviors of basilar membrane in human cochlea, comprising:
providing a processor configured to:
providing a first signal representing said speech or said audio to the plurality of filters, wherein said first signal is a digital signal; and
filtering the first signal with the plurality of filters, each of the plurality of filters producing one of a plurality of filtered signals;
wherein
each of the plurality of filters is formed from a base function via a dilation and a shift in time, wherein the base function is a digital filter representing an impulse response at a point on the basilar membrane to a corresponding frequency and the each of the plurality of filters is a digital filter formed from the base function via a dilation and a shift in time;
wherein the first signal is filtered according to:
T
[
a
i
,
b
]
=
∑
n
=
0
N
f
[
n
]
1
a
i
ψ
[
n
-
b
a
i
]
,
and wherein for the shift of b and a dilation of a i , the each of the plurality of digital filters is determined according to:
ψ
a
i
,
b
(
n
)
=
1
a
i
(
n
-
b
a
i
)
α
exp
[
-
2
π
f
L
β
(
n
-
b
a
)
]
cos
[
2
π
f
L
(
n
-
b
a
i
)
+
θ
]
u
(
n
)
,
wherein n=1, . . . , N, are discrete time instances, u(n) is a step function, and α, β, θ, and ƒ L are constants;
and wherein the impulse response of the base function at a particular frequency band is similar to an impulse response of the basilar membrane at a corresponding frequency band in the human cochlea when:
value of α>0,
value of β>0,
value of θ is selected to equate
∫
-
∞
∞
ψ
ai
,
b
(
n
)
ⅆ
n
=
0
,
u(n) =1 for n >=0, and
a>0, b>=0, ƒ L >0.
2. The method of claim 1 , wherein each of the plurality of digital filters has a frequency response to which a central frequency is associated.
3. The method of claim 2 , wherein central frequencies to the plurality of digital filters are distributed according to one of a Bark scale, an equivalent rectangular bandwidth (ERB) scale, and a log scale.
4. The method of claim 1 , wherein the α is in a range of 1 to 4 and the β is in a range of 0.01 to 0.5.
5. The method of claim 1 , wherein the α is approximately equal to a value of 3 and the β is approximately equal to a value of 0.035.
6. The method of claim 1 , further comprising:
processing the plurality of the filtered signals according a set of pre-determined rules to produce a plurality of processed signals; and
synthesizing a second speech signal based on a summation of products of each of the processed the plurality of the filtered signals and each of the plurality of filters.
7. The method of claim 6 , wherein the processing includes at least one of a reduction of noise in the speech or audio, a change of magnitude of the digital signal, and a change of number of times frames in the digital signal at an output of a cochlear transformation (CT).
8. The method of claim 6 , wherein the synthesizing the second speech signal is based on:
f
_
[
n
]
=
1
C
∑
a
i
=
a
1
a
k
∑
b
=
1
N
1
a
i
T
_
[
a
i
,
b
]
ψ
[
n
-
b
a
i
]
wherein 0≦n≦N, a 1 ≦a i ≦a k , and 1≦b≦N, T[a i ,b] represents the plurality of the filtered signals with the shift of b and a dilation of a i , and
ψ
[
n
-
b
a
i
]
.
9. The method of claim 1 , further comprising:
extracting at least one feature from the filtered signal for a speech recognition application.
10. The method of claim 9 , wherein the at least one feature includes CT cepstral coefficients.
11. The method of claim 1 , wherein the each of the plurality of filters is made from analog components.
12. The method of claim 1 , wherein the α is approximately equal to a value of 3 and the βis approximately equal to a value of 0.15.
13. A non-transitory machine-readable storage media having stored thereon instructions adapted to be executed by a processor to perform a method for processing signals representing speech or audio via a plurality of filters that approximate behaviors of human cochlea comprising:
providing a first signal representing speech to the plurality of filters, wherein said first signal is a digital signal;
filtering the first signal with the plurality of filters, each of the plurality of filters producing one of a plurality of filtered signals;
wherein
each of the plurality of filters is formed from a base function via a dilation and a shift in time, wherein the base function is a digital filter representing an impulse response at a point on the basilar membrane to a corresponding frequency and the each of the plurality of filters is a digital filter formed from the base function via a dilation and a shift in time;
wherein the signal is filtered according to:
T
[
a
i
,
b
]
=
∑
n
=
0
N
f
[
n
]
1
a
i
ψ
[
n
-
b
a
i
]
,
and wherein for the shift of b and a dilation of the each of the plurality of digital filters is determined according to:
ψ
a
i
,
b
(
n
)
=
1
a
i
(
n
-
b
a
i
)
α
exp
[
-
2
π
f
L
β
(
n
-
b
a
)
]
cos
[
2
π
f
L
(
n
-
b
a
i
)
+
θ
]
u
(
n
)
,
wherein n=1, . . . , N, are discrete time instances, u(n) is a step function, and α, β, θ, and ƒ L are constants;
and wherein the impulse response of the base function at a particular frequency band is similar to an impulse response of the basilar membrane at a corresponding frequency band in the human cochlea when:
value of α>0,
value of β>0,
value of θ is selected to equate
∫
-
∞
∞
ψ
ai
,
b
(
n
)
ⅆ
n
=
0
,
u(n)=1 for n>=0, and
a>0, b>=0, ƒ L >0.
14. A system for processing signals representing speech or audio via a plurality of filters that approximate behaviors of human cochlea, comprising:
a plurality of filters; and
a processor configured to:
providing a first signal representing speech to the plurality of filters, wherein said first signal is a digital signal;
filtering the first signal with the plurality of filters, each of the plurality of filters producing one of a plurality of filtered signals;
wherein
each of the plurality of filters is formed from a base function via a dilation and a shift in time, wherein the base function is a digital filter representing an impulse response at a point on the basilar membrane to a corresponding frequency and the each of the plurality of filters is a digital filter formed from the base function via a dilation and a shift in time;
wherein the signal is filtered according to:
T
[
a
i
,
b
]
=
∑
n
=
0
N
f
[
n
]
1
a
i
ψ
[
n
-
b
a
i
]
,
and wherein for the shift of b and a dilation of the each of the plurality of digital filters is determined according to:
ψ
a
i
,
b
(
n
)
=
1
a
i
(
n
-
b
a
i
)
α
exp
[
-
2
π
f
L
β
(
n
-
b
a
)
]
cos
[
2
π
f
L
(
n
-
b
a
i
)
+
θ
]
u
(
n
)
,
wherein n=1, . . . , N, are discrete time instances, u(n) is a step function, and α, β, θ, and ƒ L are constants;
and wherein the impulse response of the base function at a particular frequency band is similar to an impulse response of the basilar membrane at a corresponding frequency band in the human cochlea when:
value of α>0,
value of β>0,
value of θ is selected to equate
∫
-
∞
∞
ψ
ai
,
b
(
n
)
ⅆ
n
=
0
,
u(n)=1 for n>=0, and
a>0, b>=0,ƒ L >0.
15. A method for processing signals representing speech or audio via a plurality of filters that approximate behaviors of basilar membrane in human cochlea, comprising:
providing a first signal representing said speech or said audio to the plurality of filters, wherein said first signal is an analog signal; and
filtering the first signal with the plurality of filters, each of the plurality of filters producing one of a plurality of filtered signals;
wherein
each of the plurality of filters is formed from a base function via a dilation and a shift in time, wherein the base function is an analog filter representing an impulse response at a point on the basilar membrane to a corresponding frequency and the each of the plurality of filters is an analog filter formed from the base function via a dilation and a shift in time;
wherein the first signal is filtered according to:
T
(
a
,
b
)
=
∫
-
∞
∞
f
(
t
)
1
a
ψ
(
t
-
b
a
)
ⅆ
t
and wherein for the shift of b and a dilation of the each of the plurality of analog filters is determined according to:
ψ
a
,
b
(
t
)
=
1
a
(
t
-
b
a
)
α
exp
[
-
2
π
f
L
β
(
t
-
b
a
)
]
cos
[
2
π
f
L
(
t
-
b
a
)
+
θ
]
u
(
t
)
wherein n=1, . . . , N, are discrete time instances, u(n) is a step function, and α, β, θ, and ƒare constants;
and wherein the impulse response of the base function at a particular frequency band is similar to an impulse response of the basilar membrane at a corresponding frequency band in the human cochlea when:
value of α>0,
value of β>0,
value of θ is selected to equate
∫
-
∞
∞
ψ
ai
,
b
(
n
)
ⅆ
n
=
0
,
u(n) =1 for n >=0, and
a >0, b >=, 0,ƒ L >0.
16. The method of claim 15 , wherein an inverse cochlear transform transforms T [a, b] to time domain according to:
f
(
t
)
=
1
C
∫
a
=
0
∞
∫
b
=
0
∞
1
a
2
T
(
a
,
b
)
ψ
a
,
b
(
t
)
ⅆ
a
ⅆ
b
.
17. The method of claim 15 , wherein:
the base function is an analog filter denoting a continuous waveform, and
the each of the plurality of filters is an analog filter formed from the base function via a dilation and shift in time.