Running Range Normalization for Voice Activity Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional voice activity detectors are unreliable when dealing with audio signals from different environments or with varying signal-to-noise ratios, as they are sensitive to the percentage of speech in the audio signal, leading to misleading results due to shifted mean values of feature normalization.
Innovation Solution
The method involves computing running estimates of minimum and maximum values of voice activity detection features and mapping them to a desired target range using asymmetrical exponential averaging, which reduces sensitivity to speech percentage and enhances robustness across different conditions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional mean-variance normalization is used to normalize voice activity detection features, then the robustness to different data sets is improved, but the sensitivity to speech percentage changes causes misleading results when the percentage of speech in online data differs from training data
Solution Approach 1:
The patent changes the normalization parameters from mean and variance (MVN) to running minimum and maximum values. This parameter transformation makes the normalization less sensitive to speech percentage changes while maintaining robustness across different data sets, resolving the contradiction between robustness and measurement precision
Solution Approach 2:
The patent introduces dynamic running estimates of minimum and maximum values that continuously adapt to the audio signal characteristics. This dynamic approach allows the normalization to automatically adjust to varying speech percentages in real-time, preventing the misleading results that occur with static mean-variance normalization
2Device complexity
If conventional voice activity detectors with fixed thresholds are used, then the simplicity of the system is maintained, but the reliability deteriorates when applied to audio signals from different environments or with varying signal-to-noise ratios
Solution Approach 1:
The patent transforms fixed thresholds into dynamic normalization parameters (running min/max estimates) that automatically adapt to different environments and signal-to-noise ratios. This maintains system simplicity while dramatically improving reliability across varying conditions
Solution Approach 2:
The normalization system performs self-adjustment by continuously computing running estimates from the audio signal itself, eliminating the need for manual threshold calibration for different environments. The system serves itself by automatically adapting to new conditions without external intervention
3Measurement precision
If neural networks are used to improve voice activity detection accuracy, then the accuracy for training data is improved, but the reliability deteriorates when applied to audio signals from different environments with different noise characteristics
Solution Approach 1:
The patent introduces feature normalization as an intermediary step between the audio signal and the neural network. This normalization intermediary transforms the input features into a standardized format that is less sensitive to environmental variations, enabling the neural network to maintain high accuracy across different environments
Solution Approach 2:
The patent changes the feature parameters through running range normalization before they are fed to the neural network. This parameter transformation makes the features more adaptable to different environments, resolving the contradiction between training accuracy and environmental adaptability
Data Source
AI summary
A “running range normalization” method includes computing running estimates of the range of values of features useful for voice activity detection (VAD) and normalizing the features by mapping them to a desired range. Running range normalization includes computation of running estimates of the minimum and maximum values of VAD features and normalizing the feature values by mapping the original range to a desired range. Smoothing coefficients are optionally selected to directionally bias a rate of change of at least one of the running estimates of the minimum and maximum values. The normalized VAD feature parameters are used to train a machine learning algorithm to detect voice activity and to use the trained machine learning algorithm to isolate or enhance the speech component of the audio data.