Perceptual harmonic cepstral coefficients as the front-end for speech recognition
View Patent ↗Pitch estimation and classification into voiced, unvoiced and transitional speech were performed by a spectro-temporal auto-correlation technique. A peak picking formula was then employed. A weighting function was then applied to the power spectrum. The harmonics weighted power spectrum underwent mel-scaled band-pass filtering, and the log-energy of the filter's output was discrete cosine transformed to produce cepstral coefficients. A within-filter cubic-root amplitude compression was applied to reduce amplitude variation without compromise of the gain invariance properties.
1. A speech recognition method to be performed on an input speech signal using a perceptual harmonic cepstral coefficient comprising:
a) processing a speech frame to obtain a short-term power spectrum;
b) performing a robust pitch estimation on said speech signal;
c) using a peak-picking formula to the power spectrum to obtain a pitch harmonic;
d) applying class-dependent harmonic weighting to a harmonic spectrum to obtain the harmonics weighted spectrum;
e) applying a mel-scaled filter to the harmonics weighted spectrum; and
f) computing a log energy output from the harmonic weighted spectrum which is transformed into cepstrum by applying discrete cosine transform.
2. The method of claim 1 wherein the processing of said speech frame is by Fast Fourier transform or Discrete Fourier transform.
3. The method of claim 1 in which the step of performing the robust pitch estimation is in accordance with the formula:
R (τ)=β· R T (τ)+(1−β)· R s (τ)
wherein β=0.5
wherein
R
T
(
τ
)
=
∑
n
=
0
N
-
τ
-
1
[
s
~
t
(
n
)
·
s
~
t
(
n
+
τ
)
]
∑
n
=
0
N
-
τ
-
1
s
~
t
2
(
n
)
·
∑
n
=
0
N
-
τ
-
1
s
~
t
2
(
n
+
τ
)
and N is the number of samples for pitch estimation,
wherein
R
S
(
τ
)
=
∫
0
π
-
ω
τ
S
~
f
(
ω
)
·
S
~
f
(
ω
+
ω
τ
)
∫
0
π
-
ω
τ
S
~
f
2
(
ω
)
∫
0
π
-
ω
τ
S
~
f
2
(
ω
+
ω
τ
)
and ω π =2π/τ, S f (ω) is the magnitude spectrum of s i (n), and {tilde over (S)} f (ω) is the zero-mean version of S f (ω).
4. The method of claim 3 wherein the step of performing the robust pitch estimation allows for classification of the speech as voiced, unvoiced, or transitional uses the spectro-temporal auto-correlation criterion R(τ), such that if R(τ)>α v , the speech frame is classified as voiced, if R(τ)<α v , the speech frame is classified as unvoiced, and if α v ≧R(τ)≧α u , the speech frame is declared transitional, wherein α v =0.8 and α u =0.5.
5. The method of claim 1 , in which the step of obtaining the harmonics weighted spectrum is in accordance with the formula:
HWS (ω)= w h ( w )· S f (ω)
wherein w h (ω)=W H or 1 and S f (ω)=the magnitude spectrum of input speech.
6. The method of claim 5 , wherein W H =100 for voiced sounds.
7. The method of claim 5 , wherein W H =10 for transitional sounds.
8. The method of claim 1 , in which the step of applying mel-scaled filters and computing the log energy output is in accordance with the formula:
{tilde over (S)} (ω)=[ S (ω)] 1/3
Ê i =log( E i ),1≦ i≦M
wherein Ŝ(ω) is the compressed spectrum and Ê i is the log energy for band-pass filter i.
9. The method of claim 1 , further comprising: prior to performing the robust pitch estimation, applying the intensity-loudness power law to the power spectrum to obtain a root-power compressed spectrum.
10. A speech recognition method performed on an input speech signal, comprising
applying to said speech signal a harmonic weighing function having the formula:
w
h
(
ω
)
=
{
max
(
1
,
ⅇ
(
H
a
-
η
)
·
γ
)
,
if
ω
≤
ω
T
is
pitchharmonic
1
,
otherwise
wherein
H
a
=
max
τ
R
(
τ
)
is the harmonic confidence, η is the harmonic confidence threshold, γ is the weight factor, R(τ) is the spectro-temporal autocorrelation criterion, and ω T is the cut-off frequency.
11. The method of claim 10 wherein ω T , η and γ are about 2.5 kHz, 0.5 and 10, respectively.
12. A speech recognition method performed on an input speech signal, comprising:
applying to said speech signal a harmonic weighing function having the formula:
w
h
(
ω
)
=
{
max
(
1
,
ⅇ
(
H
a
-
η
)
·
γ
)
,
if
ω
≤
ω
T
is
pitchharmonic
1
,
otherwise
wherein
H
a
=
max
τ
R
(
τ
)
is the harmonic confidence, η is the harmonic confidence threshold, γ is the weight factor, R(τ) is the spectro-temporal autocorrelation criterion, and ω T is the cut-off frequency.
13. The method of claim 12 wherein η and γ 0.5 and 10, respectively.
14. The method of claim 12 wherein ω T is 4000π for transitional sounds.
15. The method of claim 12 wherein ω T is obtained for voiced sounds in accordance with the formula:
ω
T
=
arg
max
ω
T
l
<
ω
<
ω
T
h
[
∫
ω
T
l
ω
V
s
(
Ω
)
ⅆ
Ω
]
/
(
ω
-
ω
T
l
)
[
∫
ω
ω
T
h
V
s
(
Ω
)
ⅆ
Ω
]
/
(
ω
T
h
-
ω
)
wherein
V
s
(
Ω
)
=
∫
Ω
-
ω
a
Ω
S
~
f
(
ω
)
S
~
f
(
ω
+
ω
0
)
ⅆ
ω
[
∫
Ω
-
ω
a
Ω
S
~
f
2
(
ω
)
ⅆ
ω
]
[
∫
Ω
Ω
+
ω
a
S
~
f
2
(
ω
)
ⅆ
ω
]
and ω T i and ω T h delimit the allowed interval for the cutoff frequency.
16. The method of claim 15 wherein ω T i is 2000π and ω T h is 6000π.