Chosen Methods of Investigation
The limited timeframe of our project meant both DTW and HMM-based approaches were impractical, requiring many hundreds more man-hours than was available. We chose to focus on achieving solid results from a more primitive algorithm, the LPC, and work on making it more robust thereafter.
We collected the several hundreds of data samples used to train the library from ourselves.
We featured-matched input and stored data using the Yule-Walker autocorrelation method, minimizing the forward prediction error in the least squares sense. This was done using Matlab’s Yule-Walker AR Estimator.
Testing the algorithm resulted in an abysmal 20-30% accuracy.
We thought to produce better base accuracy with an algorithm of our own making. Our final results are based upon the following algorithm outlined:
- Convolution-based segmentation
- Feature extraction of formants via nonlinear power filter
- Display filtered spectrum on a discrete, weighted scatter plot
- Trace out contours of the maximum-likelihood Gaussian Mixture Model (GMM) using a maximum-likelihood GMM estimator
- Construct a standardized GMM parameter library for each number
- Find the GMM matching the input with a maximum-likelihood fit