Deep Learning | Speaker Indentification
- 1. DLL for Speaker Identification Presentation I Sai Kiran Kadam Description: Apply Deep Learning algorithms to Speaker Identification/Authentication for Cyber Security and IoT1
- 2. Recent Advances in Deep Learning for Speech Recognition at Microsoft Li Deng, Jinyu Li, Jui-Ting Huang, Kaisheng Yao, Dong Yu, Frank Seide, Michael Seltzer, GeojjZweig, Xiaodong He, Jason Williams, Yifan Gong, and Alex Acero This paper emphasizes the importance of Deep Learning over the traditional methods of speech recognition. DL with MFCC (Cepstral Analysis) demonstrated lower speech recognition and word rate errors (WER) when compared to GMM – HMM model. Higher the languages, to train DNN architecture, the lower the WER – 3.5% fewer than the DNN with a single language. DNN is a universal learner which handles heterogeneous data from different speech sources and languages as hidden layers outperform other methods. Based on research and experimental results, DNN outperforms the GMM systems. Questions: What equipment/hardware was used during training in various experiments ? How were the WER’s tracked and compared? How to implement DL Algorithms with different DNN’s? 2
- 3. Unsupervised feature learning for audio classification using convolutional deep belief networks Honglak Lee, Yan Largman, Peter Pham, Andrew Y. Ng Apply Convolutional Deep Belief Nets, CDBN is a CRBM, to unlabeled TIMIT audio data to show the learned features correspond to phonemes. (The spectral features spectrogram) CDBN after PCA of TIMIT to lower the dimensions of audio data. CDBN to Speaker Recognition CDBN with more layers outperformed RAW, MFCC Test Accuracy: CDBN(L1+L2) > CDBN-L1 > CDBN-L2 > MFCC > RAW for Speech, Music CDBN achieved higher performance on multiple audio recognition tasks. Challenge: Apply CDBN to larger data sets. Questions: What Math & Statistics concepts are used to derive the Energy equations of visible and hidden layers? Do we need to focus on Math? Does Visualization of Phonemes involve Signal Processing? 3
- 4. An Extensible Speaker Identification Sidekit In Python Anthony Larcher, Kong Aik Lee, Sylvain Meignier SIDEKIT – SpeakerIDEntification toolKIT is 100% Python tested on several platforms under Python 2.7 and > 3.4. Developed with minimum dependencies to external modules. Provides end-to-end tool-chain plus state-of-the-art methods for SIDE Easy to install, implement algorithms, enable large datasets, fast computation, compatibility with existing tools, end-to-end speaker recognition, etc. SIDEKIT is better than other existing tools like ALIZE (C++), Kaldi (C++), Matlab-SR, etc. Has tools to Evaluate Equal error Rates, Decision Cost Function, and plot Detection Error Trade-off Curves. Uses GMM-HMM based approaches. Questions: Can SIDEKIT be designed based on DNN methods? Can it be used/integrated for/with User Authentication using Keystrokes Dynamics? 4
- 5. Keystroke Dynamics User Authentication Based on Gaussian Mixture Model and Deep Belief Nets Yunbin Deng and Yu Zhong Keystroke dynamics, a behavioral biometric, offers many advancements in User Authentication depending on habitual patterns – Digraphs, Trigraphs. Uses four approaches for classification and training, namely, Statistical method based on distances, Neural Nets, Statistical ML and other algorithms. Discusses implementation of Keystroke dynamics using a 32 mixture GMM-UBM and DBN methods to reasonably represent an imposters data. DBN outperforms the GMM and GMM- UBM methods by significantly reducing the EER to 3.5%, 58% relative ERR.(static text) Highly accurate authentication is achievable by DBN Questions: Similar methodology be applied to Speaker recognition? By identifying different variations in a persons voice or speech signal? 5
- 6. Effects Of Equipment Variations On Speaker Recognition Error Rates Clark D. Shaver MS Thesis Clarks work focused on the effect of different microphone gains and variabilities on Error Rates for Speaker Recognition. He used the traditional Gaussian method for evaluating the microphones in the Recognition task. The DNN or DBN method can be implemented in the algorithm, Clark worked on. Depending on the previous slides/papers, if DBN/DNN is applied to his work, the accuracy in Error Rates of different microphones can be improved significantly. SVM can also be applied for classification purpose in Speaker Recognition. 6
- 7. Related Research Papers on DL for Speaker Recognition Fred Richardson, Douglas Reynolds, and Najim Dehak, “Deep Neural Network approaches to Speaker and Language Recognition,” IEEE, Oct 2015. Rafael G. C. P. Pinto, Hardy L C. P. Pinto and Luiz P. Ca1oba,” Using Neural Networks for Automatic Speaker Recognition: A Practical Approach,” IEEE 1996. Sreenivas Sremath Tirumala, Seyed Reza Shahamiri, “A review on Deep Learning approaches in Speaker Identification,” ICSP, Nov 2016 Alex Fandrianto, Ashley Jin, and Aman Neelappa, “Speaker Recognition Using Deep Belief Networks,” Fall 2012, Stanford University. Muhammad Muneeb Saleem,“Deep Learning For Speech Classification And Speaker Recognition,” Thesis, UT Dallas, Dec 2014. 7