Digital Signal Processing Reference
In-Depth Information
Frame Energy
Voiced speech usually has a higher energy than unvoiced speech. However,
the actual value of the energy in each frame also depends on the dynamic
range of the signal. Therefore a more useful measure is to have a comparison
of current frame energy with the tracked maximum and minimum energies.
The voiced speech should ideally be closer to the maximum track energy and
unvoiced speech should be closer to the minimum track energy (excluding
silences). The maximum track energy must go up quickly and come down
slowly and the minimum tracked energy must come down quickly and go
up slowly.
αE
max
(n
−
1
)
+
(
1
−
α)E
0
;
if E
0
> E
max
(n
−
1
)
E
max
(n)
=
(6.47)
γ E
max
(n
−
1
)
+
(
1
−
γ)E
0
;
otherwise
where
E
0
is the current frame energy and
E
max
(n
−
1
)
is the previously tracked
maximum energy. Typically
α
=
0
.
5and
γ
=
0
.
98 enables the maximum
energy to go up fast and come down slowly.
ζE
min
(n
−
+
−
;
−
1
)
(
1
ζ)E
0
if E
0
< E
min
(n
1
)
E
min
(n)
=
(6.48)
−
+
−
;
βE
min
(n
1
)
(
1
β)E
0
otherwise
where
E
min
(n)
and
E
min
(n
1
)
are the current and previously tracked minimum
energies. Typical values of
ζ
−
0
.
98 are selected so that the
minimum energy can come down fast and go up slowly. In addition to the
above tracked maximum and minimum energies, the average energy of the
speech signal may also be tracked by,
=
0
.
55 and
β
=
E
av
(n)
=
0
.
75
E
av
(n
−
1
)
+
0
.
25
E
0
(6.49)
The current frame energy, tracked average energy and tracked minimum
energy will be low in the unvoiced regions. In the voiced regions, on the other
hand, current frame energy will be close to the tracked maximum. A speech
waveform with its corresponding maximum, minimum and average tracked
energies, and the frame energy are shown in Figure 6.28. The following piece
of logic can be used to indicate voiced or unvoiced,
if ((E
0
+
th
1
> E
max
)
||
(E
0
> E
ave
))
voiced
else if (E
0
< th
2
Fe
=
+
E
min
)
(6.50)
Fe
=
unvoiced
else
Fe
=
notsure
where
th1
and
th2
are tuning tolerance thresholds.
Search WWH ::
Custom Search