Information Technology Reference
In-Depth Information
system will try to identify references, the speaker verification process follows,
which consists in the speaker confirming his/her identity. This process stage is of
great importance, as without speaker verification, an identification could not be
completed even if the person about whom the system is certain that he/she is the
speaker searched for is correctly recognised. After the verification stage is com-
pleted, the analysing system begins the second stage, referred to as 'speaker iden-
tification'. This stage identifies as such which of the speakers known to the system
is speaking. Thus this is the stage at which the entire recognition process can fin-
ish. If the speaker has been correctly linked to a person included in the system ref-
erence set (called the system base), the recognition process can be judged fully
successful, but this success is not always possible as the set constituting the data-
base may be too small and may not contain the correct references to the speaker
analysed at the given moment. Of course, disruptions may occur in the entire
analysis process or the analysing hardware may fail, and as a result the recognition
process ends in failure. However, such situations occur relatively rarely and they
do not paralyse the analyses conducted to any significant extent.
Even though the speaker analysis process does not lead to analysing the con-
tents of the spoken text, it can depend on the type of text pronounced. In the case
of speaker analysis independent of the spoken content we can claim that this
analysis should proceed correctly and what is more, should end in success, namely
speaker recognition, for any statement of this speaker. Yet it is worth adding that
this type of analysis is used in situations in which we are convinced that cooperat-
ing with the speaker will not produce any benefits or when the speaker does not
want to cooperate with us (as the people conducting the analysis). In other cases,
methods of speaker recognition that may depend to some extent on the content
said out loud are very often used. These are cases where the speaker is suspected
of using or pronouncing a password or a fragment thereof, uses numbers coded in
the system or pronounces a text provided by the analysing system. In these cases,
the speaker recognition is said to be dependent on the content pronounced by the
speaker.
Thus speaker recognition is determined by the presence of a certain type of a
link between the specific speaker and a certain pattern defining this speaker. If the
system contains a pattern defining the speaker, then in the recognition process the
system will assign the speaker to his/her reference model. If the opposite is true,
then the system may indicate the reference pattern closest to the speaker, or if the
defined number of consistent features is not achieved, it may signal that the recog-
nition analysis failed. The case in which the speaker corresponds ideally to the
pattern with which he/she has been associated and to which he/she is assigned is
called 'speaker identification from a closed set of speaker models', whereas if a
certain number of consistent features is analysed because there is no full compli-
ance, this is referred to as 'identifying the speaker from an open set of speaker
models'.
The speaker recognition process, regardless of the type of the speaker and of
the analysis conducted, consists of three fundamental stages, namely: the prelimi-
nary processing of the speech signal, the selecting features significant for the
