6.1. Analysis toolkit
145
6.1.6.1
Instantaneous LSP analysis
Although there are many possible methods of analysis, in this section we will consider
just three features which can be used for individual analysis frames of speech. Later,
in Section 6.1.6.2, we will then extend the analysis to the evolution of the measures
through time. In the following equations, subscript i refers to one of the p LSPs
ω
i
which
represent speech in analysis frame n.
The overall shift in LSP location from one frame to the next, referred to as the gross
shift, Shift, indicates predominant spectral movements between frames:
Shift
[n] =
p
i
=1
ω
i
[n]
−
p
i
=1
ω
i
[n + 1]
.
(6.5)
In Matlab this would be:
function [shift] = lsp_shift(w1,w2)
shift=sum(w1) - sum(w2);
The mean frequency of all LSPs representing an analysis frame is also a useful measure
of frequency bias, Bias:
Bias
[n] =
1
p
p
i
=1
ω
i
[n]
(6.6)
or in Matlab:
function [bias] = lsp_bias(w)
bias=sum(w)/length(w);
It is also possible to specify a nominal LSP positioning, and calculate a deviation Dev
between this reference LSP positioning
¯ω
i
and the LSPs under analysis, raised to an
arbitrary power
β. This is measured as follows:
Dev
[n] =
p
i
=1
(ω
i
[n] − ¯ω
i
)
β
.
(6.7)
A suitable Matlab function for calculating the deviation is shown below:
function [dev] = lsp_dev(w,bar_w,b)
dev=sum( (w-bar_w).ˆb);
where
¯ω
i
would often be a set of nominally positioned LSPs equally spaced across the
frequency domain:
¯ω
i
= iπ/(p + 1)
for
i
= 1, . . . , p.
(6.8)
146
Do'stlaringiz bilan baham: |