Download presentation
Presentation is loading. Please wait.
Published byBrandon Lane Modified over 8 years ago
1
Madhulika Pannuri Intelligent Electronic Systems Human and Systems Engineering Department of Electrical and Computer Engineering Correlation Dimension on speech (sustained phones)
2
Page 1 of 13 Madhulika Pannuri Analysis setup This analysis includes estimates of dimension for varying initial embedding dimension and data size. Variation in dimension estimates with SNR of the signal was studied. The analysis was performed on 3 vowels, 2 nasals and 2 fricatives. Statistical modelling with male model, female model and cross speaker model was performed.
3
Page 2 of 13 Madhulika Pannuri Varying initial embedding dimension (Vowels)
4
Page 3 of 13 Madhulika Pannuri Varying the initial embedding dimension (nasals)
5
Page 4 of 13 Madhulika Pannuri Varing the initial embedding dimension (fricatives)
6
Page 5 of 13 Madhulika Pannuri Varying data size (Vowels)
7
Page 6 of 13 Madhulika Pannuri Varying data size (nasals)
8
Page 7 of 13 Madhulika Pannuri Varying the data size (fricatives)
9
Page 8 of 13 Madhulika Pannuri Varying SNR (Vowels)
10
Page 9 of 13 Madhulika Pannuri Varying SNR (Nasals)
11
Page 10 of 13 Madhulika Pannuri Varying SNR (Fricatives)
12
Page 11 of 13 Madhulika Pannuri KL Divergence measures for various phonemes
13
Page 12 of 13 Madhulika Pannuri Statistics of dimension
14
Page 13 of 13 Madhulika Pannuri Aurora Plans: - Analyze speaker variability for sustained phones. - Add other nonlinear invariants to feature vector and analyze the improvement in performance for AURORA. Mixtures% WER 1 17.7 % 16 10.1 %
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.