Boosted Deep Neural Networ ks and Multi-resolution

Typical nonparametric models:
support vector machines (Enqing et al., 2002) sparse coding spectral clustering
Typical parametric models:
Gaussian models, Gaussian mixture models recursive neural networks deep neural networks (Zhang and Wu, 2013)
10
2 The cochleagram feature is calculated from ∑K k =1 sc ,k where K —frame length, sc ,k —value of the c th channel and k th sample of the frame.
Xiao-Lei Zhang BoostDNN and MRCG for VAD
1 Speech hit rate
Optimization direction
ROC curve
Optimal operating point
0
0 False alarm rate
1
3
Xiao-Lei Zhang
BoostDNN and MRCG for VAD
Intro. Related Work bDNN MRCG Experiments Conclus. Ack
Boosted Deep Neural Networks and Multi-resolution Cochleagram Features for Voice Activity Detection
Xiao-Lei Zhang
∗ Department of Electronic Engineering, Tsinghua University, Beijing, China. Department of Computer Science & Engineering, The Ohio State University, USA (e-mail: huoshan6@).
Frame length = 200 ms; Frame shift = 10 ms
d
CG1
CG3
Smoothing each unit in a 11x11 square window
Smoothing each unit in a 23x23 square window
Frame length
64-D feature
8 September 2014
1
Intro. Related Work bDNN MRCG Experiments Conclus. Ack
Contents
1
Introduction Related Work Boosted Deep Neural Networks Multiresolution Cochleagram Features Experiments Conclusions Acknowledgments
Typical features: DFT, MFCC, Pitch, etc.
Challenges:
How to improve the performance in extremely low SNRs and difficult environments? How to generalize the models to unseen noise types?
Training phase of boosted deep neural networks (bDNN)
a
b
DNN
b
DNN
Figure: Training phase on a corpus {xn , yn }N n=1 . (a) Expanding xn to T T T T xn = [xT n−W , xn−W +1 , . . . , xn , . . . , xn+W −1 , xn+W ] given the half-window size W . (b) Expanding the ground truth label yn to a new (2W + 1)-dimensional target vector yn = [yn−W , yn−W +1 , . . . , yn , . . . , yn+W −1 , yn+W ]T 7
5
Xiao-Lei Zhang BD
Intro. Related Work bDNN MRCG Experiments Conclus. Ack
Related Work
Deep neural networks (DNN): DNNs are multilayer perceptrons with more than one hidden layer.
9
Intro. Related Work bDNN MRCG Experiments Conclus. Ack
Experimental settings
Data sets: Clean speech of AURORA4 mixed with Babble and factory noises from NOISEX-92 at an SNR of [-5 0 5] dB Evaluation metrics: ROC curve and its corresponding AUC Competitive methods:
Bootstrap resamplings:
Bootstrap (Efron, 1979) Key idea: integrating the results of multiple subsets (called bootstrap samples) for a smaller variance than any of the results. Ensemble learning (Schapire, 1990), Key ideas: (i) meaningful weak classifiers, and (ii) diversity between the classifiers. Methods (Dietterich, 2000):
CG4
Speech signal
64-channel gammatone filter
(Frequency range: [80, 5000] Hz)
64-D feature
Time
Frame length
A Delta feature is calculated by ∆xn =
(xn+1 −xn−1 )+2(xn+2 −xn−2 )
Manipulating training examples, input features, hyperparameters, and/or output targets.
6
Xiao-Lei Zhang BoostDNN and MRCG for VAD
Intro. Related Work bDNN MRCG Experiments Conclus. Ack
Brief overview of VADs
Energy based VADs
Typical methods:
energy in the time domain, zero-crossing rate (Rabiner, 1976), pitch detection,
Typical systems:
4
Xiao-Lei Zhang BoostDNN and MRCG for VAD
Intro. Related Work bDNN MRCG Experiments Conclus. Ack
Brief overview of VADs (2)
Machine Learning based VADs (i) view VAD as a binary-class classification problem, and (ii) train various classification models.
64-D feature
64-D feature
64-D feature
Channels
CG2
b c
256-D MRCG feature
256-D Delta feature
256-D DeltaDelta feature Calculating the energy of each frame in each channel
Multiresolution cochleagram (MRCG) (Chen et al., 2014)
a
Time domain noisy speech 64-channel cochleagram:
Frame length = 20 ms; Frame shift = 10 ms
64-channel cochleagram:
G.729B (1998), ETSI Advanced Front-End (ES 202 050, 2004).
A useful technique: Long term smoothing (e.g. 200 ms).
Statistical signal processing based VADs
Typical nonlinear hidden units: sigmoid function, tanh function, and rectified linear function. Typical training methods: backpropagation, pretraining.