| Title | : | Language Identification using SVD on UBM-GMM supervectors |
| Speaker | : | Manish Balchand Jain (IITM) |
| Details | : | Wed, 26 Jul, 2017 3:00 PM @ A M Turing Hall |
| Abstract: | : | The state-of-the-art approach to Spoken Language Identification (LID) is based on the Total Variability Space (TVS). TVS is a paradigm that was effectively used for Speaker Verification. Unlike speaker verification, the number of classes is small. Moreover, TVS approach is extremely computationally intensive. The objective of this research work is to propose a simpler approach using SVD instead of TVS. Similar to the TVS approach the UBM-GMM is first built, and utterance-wise adapted supervectors are stacked to form an supervector matrix. Dimensionality reduction is performed on the supervector matrix using SVD. Top L singular values are chosen using two different criteria based on reconstruction error and energy based rule. The singular vectors corresponding to these L singular values are called projection vectors. These L projection vectors span the language space. The supervector matrix is projected along these projection vectors directions. During training, an SVM-based language classifier is trained on these projected vectors. During testing, each test utterance is aligned with the UBM and projected along the same L directions. The projected test vector is classified using the SVM classifier. This approach is evaluated on CallFriend corpus using various utterance durations. The proposed SVD based approach is compared with the UBM-GMM system and the i-vector system using different test cases. The proposed approach outperforms UBM-GMM system and TVS system in all the cases. The SVD system outperforms the best-performing i-vector LID system with an absolute improvement of 8.4%. Further, the i-vectors of TVS system are analyzed for the orthogonality property. Projecting i-vectors in orthogonal directions improves the i-vectors system performance by 6.4%.​ Keywords: Language Identification, SVD, i-vectors, Supervectors, Appropriate dimension. |
