Cross Comparison Of Speakers:
Recognition performance has been also computed by comparing
speech samples of male and female speakers. This is shown in Table4.
Comparison
Accuracy
(% Recognition)
Male-Male (38*38)
69.86
Male-Female (38*62)
30.28
Female-Female (62*62)
62.44
Table-4
It may be observed that best performance is of male to male speaker and poor performance is
observed on
comparison of male to female speaker.
Comparison o f H e a d -held M i c r o p h o n e and Mobile Phone Database:
As recording
has been done through different channels, therefore comparison also has been made for
speech of different
microphones. Based on different channels, accuracies on comparison of 100 speakers
have been shown in
table5:
Comparison
Accuracy (%)
Headheld-Headheld
72.20
Headheld-mobile phone
44.20
Mobile - Mobile
69.42
Table-5
Conclusions:
x
The recognition performance decreases as the number of speakers increases.
x
Performance of male speakers is better than the female speakers.
x
The p e r fo r ma nc e i nc r e a s e s a s t he n u mb e r o f utterances in training samples increases.
x
Performance of head-held microphone is better than mobile phone.
x
On c o m p a r i s o n w i t h t h e earlier experimental results conducted by different researchers, it is
concluded that here the word level recognition give about 75% accuracy which is better than the
earlier experiments
Agrawal et al.
Proceedings of Meetings on Acoustics, Vol. 19, 060019 (2013) Page 5
Future Work:
Speaker
identification results
have been evaluated for 100 speakers under clean
and noisy background conditions. For our research work we restricted ourselves to early integration
technique, in future the same work can
be implemented for other combination of integration
techniques.
In this Experiment which is conducted for 100 speakers is based on monophones hmm models. The
triphones based models work far better than monophones models. Also at the same time work is to be
done for text-independent speaker identification.
Same work can be carried out using different classifiers such as Artificial neural
network
(ANN),GMM,RASTA which can also be used and their performance can be compared. Since HTK is a
statistical based approach in which variations in speech are modeled statistically using automatic
learning
procedure, to avoid this problem, same work can be done using ANN.
Do'stlaringiz bilan baham: