SLIDE 1
E9 205 Machine Learning for Signal Procesing
20-11-2019
Deep Learning for Audio and Vision
Speech Recognition
Noise Channel Automatic Speech Systems
Courtesy – Google Images
E9 205 Machine Learning for Signal Procesing Deep Learning for Audio - - PowerPoint PPT Presentation
E9 205 Machine Learning for Signal Procesing Deep Learning for Audio - - PowerPoint PPT Presentation
E9 205 Machine Learning for Signal Procesing Deep Learning for Audio and Vision 20-11-2019 Speech Recognition Noise Channel Automatic Speech Systems Courtesy Google Images Signal Modeling Short-term spectra integrated in mel frequency
SLIDE 2
SLIDE 3
▪
Short-term spectra integrated in mel frequency bands followed by log compression + DCT – mel frequency cepstral coefficients (MFCC) [Davis and Mermelstein, 1979].
Signal Modeling
Short-term Spectrum Integration + Log + DCT 25ms
SLIDE 4
Mel Frequency Cepstral Coefficients
▪
MFCC processing repeated for every short-term frame yielding a sequence of features. Typically 25ms frames with 10ms hop in time.
SLIDE 5
- Map the features to phone class. Using phone labelled data.
Speech Recognition
/w/ /^/ /n/ w - |^| n Triphone Classes
- Classical machine learning - train a classifier on speech training data
that maps to the target phoneme class.
SLIDE 6
Back to Speech Recognition
Mapping Speech Features to Phonemes
SLIDE 7
Back to Speech Recognition
Mapping Speech Features to Phonemes to words
Language Model [Dictionary of Words Pronunciation Model Word Syntax] Decoded Text
SLIDE 8
State of Progress
2018 5.3%
Claims of human parity using BLSTM based Models !!!
SLIDE 9
Moving to End-to-End
Audio Features Text Output
SLIDE 10
Image Processing
SLIDE 11
Visual Graphics Group Network
SLIDE 12
ImageNet Task
1000 images in each of 1000 categories. In all, there are roughly 1.2 million training images, 50,000 validation images, and 150,000 testing
- images. ImageNet consists of variable-resolution images. Therefore, the
images have been down-sampled to a fixed resolution of 224×224.
SLIDE 13
Can we go deeper
SLIDE 14
SLIDE 15
Residual Blocks
SLIDE 16
Deep Networks with Residual Blocks
SLIDE 17
Deep Networks with Residual Blocks
SLIDE 18
Results with ResNet
SLIDE 19
Image Segmentation
SLIDE 20
SLIDE 21
The Problem of Segmentation
SLIDE 22
SegNet Architecture
SLIDE 23
Results from Segnet
SLIDE 24
SLIDE 25
U-net
SLIDE 26
Summary of the Course
SLIDE 27
Distribution Pie Chart
55% 45% Generative Modeling and Dimensionality Reduction Discriminative Modeling
SLIDE 28
Generative Modeling and Dimensionality Reduction
15% 15% 8% 31% 15% 15%
Feature Processing PCA/LDA Gaussian and GMM NMF Linear and Logistic Regression kernel methods
SLIDE 29
Discriminative Modeling
11% 17% 6% 6% 6% 6% 11% 11% 11% 17%
SVM Neural Networks Improving Learning Improving Generalization Deep Networks
- Conv. Networks
RNNs Understanding DNNs Deep Generative Modeling Applications
SLIDE 30
When we started …
SLIDE 31
Dates of Various Rituals
❖ 5 Assignments spread over 3 months (roughly one assignment every two
weeks).
❖ September 1st week - project topic announcements. ❖ September 3rd week - 1st Midterm ❖ September 4th week - project topic and team finalization and proposal
- submission. [1 and 2 person teams].
❖ October 1st week - Project Proposal ❖ October 3rd week - 2nd MidTerm ❖ November 1st week - Project MidTerm Presentations. ❖ December 1st week - Final Exams ❖ December 2nd week - Project Final Presentations.
SLIDE 32
Content Delivery
Implementation and Understanding
Theory and Mathematical Foundation Intuition and Analysis