Wavelet Features For Speech Recognition.¶

I've been partipating in the TensorFlow speech recognition challenge.

https://www.kaggle.com/c/tensorflow-speech-recognition-challenge

It seems like the most common approach is to begin by turning the audio into a spectrogram and then feeding that into a 2D CNN. One trouble with spectrograms is that you have to trade off resolution in frequency for resolution in time and vice versa. In principle you can get higher resolution in time for higher frequencies than you can for lower frequencies but when you pick an input length for your short time fourier transform you lose temporal resolution much below the window length.

Wavelets are one possible way around this limitation

Intermezzo: Sparsely Observed Data¶

In the post on using PCA for data imputation we used a weight for each of our data points. By assigning a weight of 0 to missing data and a weight of 1 to the rest of our data we managed to be able to get a reasonably good approximation to what we would find using PCA on the dataset without any data missing.

This is fine when evaluating a dense model for our data matrix is not too much computational overhead. However when our input data are sparsely observed, that is to say most of our data consists of missing values then evaluating the model densely is a tremendous waste of computational resources.

Lattice SVM¶

A support vector machine (SVM) is a classifier that attempts to find a maximum margin linear separator for different classes in a very high dimensional implicit feature space. The feature space is usually not explicitly calculated but is instead accessed via a kernel function which provides the effective dot product in the feature space, this has the advantage that we can deal with very large implicit feature spaces this way. In fact the dimensionality of the implicit feature space of most commonly used SVM variants is usually quoted as being infinite, for example the Gaussian kernel is one example. But the high effective feature dimensionality still comes with a high computational cost, we must somehow deal with an N by N matrix of similarities relating all of our training points to each other (the matrix of kernelized "feature dot products").

Asymptotic Labs (Posts about machine learning)

SKM Embedding of MNIST

Wavelet Spectrograms for Speech Recognition

Wavelet Features For Speech Recognition.¶

Low Rank Approximation On Sparsely Observed Data

Intermezzo: Sparsely Observed Data¶

Learning TensorFlow via a 3D printing Project

Lattice SVM

Lattice SVM¶