Score Fusion for Speaker Identification using MFCC and ICMC Features
Joshi, Sonal
Score Fusion for Speaker Identification using MFCC and ICMC Features by Sonal Joshi - IIT Jodhpur Department of Electrical Engineering 2017 - xi,58p. HB
"The task of Speaker Identification (SID) or Speaker Recognition is to recognise the person
from a given speech utterance. That means to answer the question, ”Whose voice is this?” An important
application of SID is in forensics to verify the identity of a suspect. Other than forensic
applications, this technology is used to improve the performance of speech recognition, automatically
adjust preferences as per personal needs like in home automation and identify the speaker in
each segment of a teleconference or newsroom discussion (Speaker Diarization).
Even though real world applications demand robustness against various possible practical
and realistic conditions, generally SID systems have poor performance when there is a mismatch.
Different recording conditions in training and testing data lead to mismatch, which can be in the
form of language mismatch, session mismatch, sensor mismatch or any combination of the above.
To improve speaker recognition performance in mismatch scenarios, score fusion of log-likelihood
scores obtained using Gaussian Mixture Model - Universal Background Model (GMM-UBM) classifier
is explored in this work.
After an initial study of commonly used features using TIMIT database, GMM-UBMs using
Mel Frequency Cepstral Coefficients (MFCC) and recently proposed Infinite impulse response
Constant Q Mel-frequency cepstral Co-efficients (ICMC) features are scored independently. This
work is motivated by the fact fusion of systems using MFCC and ICMC at the score level will lead to
performance gain as both the features have complementary information. Experimental results, obtained
using IITG Multivariability Speaker Recognition Phase-I and Phase-II Database, prove that
the fusion results outperform the independently scored results by a significant margin for all mismatches.
Reported average relative improvements in identification accuracy over baseline MFCC
in percent for 128 mixture gaussian are 1.99% for language mismatch, 4.56% for session mismatch,
5.38% for language and session mismatch, 204.54% for sensor mismatch, and 175.3% for sensor
and session mismatch. Experimental results are also obtained using IITG Multivariability Speaker
Recognition Phase-III which is a truly conversational data collected over phone call.
i"
MFCC and ICMC Features
MTech Theses
Department of Electrical Engineering
621.36 / J837S
Score Fusion for Speaker Identification using MFCC and ICMC Features by Sonal Joshi - IIT Jodhpur Department of Electrical Engineering 2017 - xi,58p. HB
"The task of Speaker Identification (SID) or Speaker Recognition is to recognise the person
from a given speech utterance. That means to answer the question, ”Whose voice is this?” An important
application of SID is in forensics to verify the identity of a suspect. Other than forensic
applications, this technology is used to improve the performance of speech recognition, automatically
adjust preferences as per personal needs like in home automation and identify the speaker in
each segment of a teleconference or newsroom discussion (Speaker Diarization).
Even though real world applications demand robustness against various possible practical
and realistic conditions, generally SID systems have poor performance when there is a mismatch.
Different recording conditions in training and testing data lead to mismatch, which can be in the
form of language mismatch, session mismatch, sensor mismatch or any combination of the above.
To improve speaker recognition performance in mismatch scenarios, score fusion of log-likelihood
scores obtained using Gaussian Mixture Model - Universal Background Model (GMM-UBM) classifier
is explored in this work.
After an initial study of commonly used features using TIMIT database, GMM-UBMs using
Mel Frequency Cepstral Coefficients (MFCC) and recently proposed Infinite impulse response
Constant Q Mel-frequency cepstral Co-efficients (ICMC) features are scored independently. This
work is motivated by the fact fusion of systems using MFCC and ICMC at the score level will lead to
performance gain as both the features have complementary information. Experimental results, obtained
using IITG Multivariability Speaker Recognition Phase-I and Phase-II Database, prove that
the fusion results outperform the independently scored results by a significant margin for all mismatches.
Reported average relative improvements in identification accuracy over baseline MFCC
in percent for 128 mixture gaussian are 1.99% for language mismatch, 4.56% for session mismatch,
5.38% for language and session mismatch, 204.54% for sensor mismatch, and 175.3% for sensor
and session mismatch. Experimental results are also obtained using IITG Multivariability Speaker
Recognition Phase-III which is a truly conversational data collected over phone call.
i"
MFCC and ICMC Features
MTech Theses
Department of Electrical Engineering
621.36 / J837S