Signal processor employing neural network trained using evolutionary feature selection

Inventors

Lee, David • Blanchard, Scott • Dodd, Nickolas

Assignees

General Dynamics Mission Systems Inc

Abstract

The evolutionary feature selection algorithm is combined with model evaluation during training to learn feature subsets that maximize speech/non-speech distribution distances. The technique enables ensembling of low-cost models over similar features subspaces increases classification accuracy and has similar computational complexity in practice. Prior to training the models, feature analysis is conducted via an evolutionary feature selection algorithm which measures fitness for each feature subset in the population by its k-fold cross validation score. PCA and LDA based eigen-features are computed for each subset and fitted with a Gaussian Mixture Model from which combinations of feature subsets with Maximum Mean Discrepancy scores are obtained. During inference, the resulting features are extracted from the input signal and given as input to the trained neural networks.

Core Innovation

The invention provides a method for using a neural network to process speech signals. It uses a corpus of data containing speech and non-speech, and a collection of feature sets where each feature set represents a different way to express a signal as a feature vector. The method performs evolutionary feature selection using a genetic algorithm to evolve from the collection of feature sets a subset collection of feature sets that maximizes a distribution distance between the speech and the non-speech.

After selecting the subset collection of feature sets, the method extracts from the corpus of data features corresponding to the subset collection of feature sets to develop training data. The training data is applied to a neural network to develop a learned model, and the approach includes performing validation of the learned model and iteratively repeating the evolutionary feature selection if the learned model does not meet a predefined stopping threshold.

The method then deploys the learned model in a processing system configured to extract from an input signal features corresponding to the subset collection of feature sets. Those extracted features are processed by the neural network based on the learned model. The disclosed system is described in the partial content as a bi-directional LSTM-RNN used for speech detection and enhancement.

Claims Coverage

The independent claim is clm-00001. It defines a complete end-to-end method including evolutionary feature selection via a genetic algorithm, training data extraction, neural network training to develop a learned model, validation with an iterative stopping threshold, and deployment in a processing system that extracts features from an input signal for neural-network processing. Dependent refinements include constructing training data using multiple speech signal processing application-specific models and validating using k-fold cross validation scores.

The claim coverage is centered on selecting feature sets with a genetic algorithm to maximize distribution distance between speech and non-speech, training a neural network using features extracted according to the selected subset, validating the learned model with an iterative stopping threshold, and deploying the learned model to process extracted features from an input signal.

Stated Advantages

Documented Applications