Bluetooth Cochlea Zhao Zijian 5130309770 Bluetooth Cochlea Android
Bluetooth Cochlea Zhao Zijian 5130309770
Bluetooth Cochlea
Android Programming • Interface of Scene Selection • Login and Register Content Audio Scene Recognition(survey) • Why we need this - Background • What others have done - Product • How we can do this - Algorithm
Android Programming Interface of Scene Selection
Android Programming Login and Register
Android Programming • Interface of Scene Selection • Login and Register Content Audio Scene Recognition(survey) • Why we need this - Background • What others have done - Product • How we can do this - Algorithm
Why we need this - Background
What others have done - Product Australia Cochlear Inc Scene Classifier America Advanced Bionics Inc Auto Sound • Broadest Input Dynamic Range Austria Medical Electronic Inc Automatic Sound Management • Broad Input Dynamic Range • Automatic Volume Control
Algorithm – Feature Extraction Stationary frequency-based Frequency extraction Mel frequency LPC coefficients Perceptual linear prediction features Linear prediction cepstral coefficients Non-stationary time-frequencybased Wigner-Ville distribution Short-time Fourier transform Continuous wavelet transform Fast discrete wavelet transform Feature Extraction
Algorithm - Classification • • • Dynamic time warping (DTW) Hidden Markov models (HMM) Learning vector quantization (LVQ) Self-organizing maps (SOM) Ergodic-HMMs Artificial neural networks (ANN) Long-term statistics Maximum likelihood estimation (MLE) Gaussian mixture models (GMM) Support vector machines (SVM) Deep neural networks
Dynamic time warping(DTW) Pattern Matching Dynamic Programming
Deep neural network • Pre-training of DNN via DBN • Fine-training of DNN Input: audio features Output: audio context class Dataset: LITIS Rouen audio dataset 2 hidden layer with 50 neurons : 80% 5 hidden layer with 500 neurons: 91. 6% 7 hidden layer with 1000 neurons: 92. 2% Reference: DEEP NEURAL NETWORKS FOR AUDIO SCENE RECOGNITION(2015)
How we can do this Dynamic time warping • Need good patterns Deep neural network • Need training dataset Environmental sounds Dataset • http: //www. cs. tut. fi/~heittolt/datasets Method • Use DNN to train a classifier • Use clustering to generate patterns, then use DTW to do pattern matching
Thanks! Any Questions?
- Slides: 14