|
« Back
Accent Identification in Speech
Voice biometrics offer major advantages over other types of
authentication techniques in terms of usability, cost, easy of
deployment and user acceptance, since it is the most non-intrusive
amongst the many biometrics that are being used. Apart from the
content of speech and identity of the user, information such as
accent, gender, age and other soft biometric traits of the individual
can also be inferred from his speech signal. This has particular
importance in situations such as 911 call handling, where gender, age
or ethnicity can be used to route the calls to specific personnel. As
a first step towards this goal, we have attempted to automatically
identify gender and accent of individuals. Human discriminate between
men and women according to the frequency of the speakers voice. Women
speak with higher fundamental frequencies as compared to men. We have
used 3 pitch estimation methods and used them in combination to
classify a speaker in 6 classes.
The system is implemented and tested
on available dataset. Overall system accuracy lies around 70%. The
accuracy can be improved a lot with larger dataset. We are currently
working on accent identification using difference in formant frequency
distribution between native and non-native speakers. We have
identified specific phoneme segments that work as accent markers in
speech and have experimentally verified that they can be used to
successfully differentiate between native and accented speech.
|
|
 Most biometrics are distinctive but not necessarily unique biometric data fusion allows one-to-one mapping between data and individuals.
|
