Content Details | NUML Online Research Repository

Title Extraction of Accent Information from Urdu Speech for Forensic Speaker recognition

Author(s) Falak Tahir

Abstract This thesis presents a new method for extraction of accent information from Urdu speech signals. Accent is used in speaker recognition system especially in forensic cases and plays a vital role in identifying people of different groups, communities and origins due to their different speaking styles. Other applications of accent are telephone banking, voice dialing, e-health and biometric authentication. This thesis focuses on only the forensic applications of the accent. Forensic detection through accent helps in criminal investigation and provides additional information such as territorial origins of the suspects. The proposed method is based on Gaussian Mixture Model-Universal Background Model (GMM-UBM) and a new Feature Mapping (FM) process. The proposed method is named as GMM-FM. The FM process maps Mel-Frequency Cepstral Coefficients (MFCC) features to higher dimensional space and improves the accent extraction and forensic speaker recognition performances of GMM-UBM. In the proposed method, GMM-UBM is used to obtain accent independent model. For this purpose the training MFCC features of the training set are processed with the proposed FM method. The processed features of all the accent categories of the training set are combined and different GMM components are computed with GMMUBM. Each GMM component is parameterized by a mean vector, mixture weight and covariance matrix. In the second step, the GMM components estimated for accent independent model are used in a Bayesian process to adapt GMM components for each accent category of the training set. Such GMM components are referred to as accent dependent GMM. To classify accent in a speech sample the log-likelihood is computed using the GMMs of both accent dependent and independent models. Then accent is predicted for the test sample based on maximizing the log-likelihood values. Experiments are performed on Urdu and Kaggle accent corpuses. The experimental results show that the proposed GMM-FM obtains on average 2.5% and 3.5% better equal error rate and accuracy than GMM-UBM, respectively. Keywords: Accent, Urdu corpus, speech signals, Gaussian components, speech features, recognition and forensic.

Type Thesis/Dissertation MS

Faculty Engineering and Computer Science

Department Computer Science

Language English

Publication Date 2019-05-30

Subject Computer Science, Machine Learning, Accent Recognition

Publisher NUML

Contributor(s) Supervisor, Dr. Sajid Saleem

Format

Identifier

Source

Relation

Coverage

Rights