Neural Network Speech Detection Using Evolutionary Feature Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional speech enhancement and voice activation detection systems face challenges in distinguishing human speech from noise, especially in non-stationary environments, due to the unpredictability of noise sources and their similarity to speech artifacts, leading to degraded speech recognizability and missed detection of speech segments.
Innovation Solution
A deep learning approach using a long short-term memory (LSTM) recurrent neural network (RNN) trained with an evolutionary feature selection process, which automates the selection of minimally redundant and maximally relevant feature subsets through a genetic algorithm, to enhance speech detection and separation from noise.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional speech enhancement systems use spectral subtraction to remove noise, then speech intelligibility may be improved, but the system fails when noise resembles speech artifacts and cannot distinguish between noise and speech
Solution Approach 1:
The patent replaces conventional mechanical signal processing methods (spectral subtraction, frame-based filtering) with a neural network-based cognitive system. The neural network learns to distinguish speech from noise through training on diverse audio data, enabling it to handle non-stationary noise environments where conventional methods fail because they cannot adapt to changing noise characteristics that resemble speech.
Solution Approach 2:
The patent transforms the input audio signal into a different parameter space using log-power spectra and other feature transformations. This parameter change allows the neural network to process audio data in a representation that highlights speech characteristics while suppressing noise, improving the system's ability to distinguish speech from non-stationary noise compared to working in the raw time or frequency domain.
2Productivity
If voice activation detection systems segment signals into frames for binary classification, then processing efficiency is improved, but speech segments may be missed when masked by noise
Solution Approach 1:
The patent replaces the conventional frame-based binary classification approach with a neural network that processes audio data through learned feature representations. This substitution maintains processing efficiency while improving reliability because the neural network can detect speech patterns even when they are partially masked by noise, avoiding false negatives that occur with threshold-based conventional systems.
Solution Approach 2:
The patent performs preliminary action by pre-training the neural network on diverse audio data including various noise conditions before deployment. This preliminary training equips the system with the ability to recognize speech patterns under different conditions, so when actual speech detection is needed, the system is already prepared to reliably distinguish speech from noise without requiring extensive real-time analysis.
3Ease of manufacture
If neural networks are trained with conventional feature sets, then training can be performed with standard data, but the network fails to discriminate between speech and noise effectively in noisy environments
Solution Approach 1:
The patent performs preliminary action by pre-processing audio data into log-power spectra and other transformed features before training. This preliminary transformation of the training data into a specialized feature space enables the neural network to learn effective speech-noise discrimination during training, rather than requiring complex manual feature engineering or post-training adjustments.
Data Source
AI summary
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.