Voice Activity Detector Using Gaussian Mixture Model Noise Correction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing voice activity detection methods in noisy environments, such as those found in maintenance work or factory automation, suffer from low accuracy due to incorrect differentiation between noise and voice activity, especially when the noise environment is unknown during the learning process.
Innovation Solution
A learning device that utilizes a synthetic neural network and a Gaussian mixture model to differentiate between noise and voice by learning bottleneck feature values, which are then used to correct the likelihood of voice activity detection, thereby improving detection accuracy in unknown noise environments.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If power-based voice activity detection is used in noisy environments, then detection sensitivity is improved, but noise is erroneously detected as voice activity reducing accuracy
Solution Approach 1:
The patent transforms the detection approach by changing from power-based parameters to spectral feature parameters (MFCC, spectral flux, zero-crossing rate). These spectral parameters better represent voice characteristics and are less susceptible to noise interference, thereby improving detection accuracy in noisy environments.
Solution Approach 2:
The patent introduces an intermediary classification mechanism using support vector machines (SVM) and artificial neural networks (ANN) that act as mediators between the raw signal features and the final voice activity detection decision. These intermediaries process spectral features to distinguish voice from noise more effectively than direct power-based detection.
2Measurement precision
If HMM with cepstrum parameters is used for voice activity detection, then noise discrimination is improved in known noise environments, but detection accuracy deteriorates in unknown noise environments
Solution Approach 1:
The patent transitions from static HMM parameters learned in specific noise environments to dynamic spectral feature extraction that adapts to any noise environment. The spectral features (MFCC, spectral flux, zero-crossing rate) are computed in real-time from the input signal, enabling the system to adapt to unknown noise conditions without requiring pre-learning of noise characteristics.
Solution Approach 2:
The patent changes the parameter representation from cepstrum-based HMM parameters to spectral feature vectors that capture instantaneous signal characteristics. This parameter transformation enables the system to generalize to unknown noise environments by relying on fundamental spectral properties rather than environment-specific statistical models.
3Measurement precision
If multiple neural networks are synthesized and Gaussian mixture models are learned, then detection accuracy in unknown noise environments is improved, but device complexity increases
Solution Approach 1:
The patent segments the complex detection task into distinct processing stages: spectral feature extraction (MFCC, spectral flux, zero-crossing rate), bottleneck feature generation through neural networks, and final classification using Gaussian mixture models. This segmentation allows each component to be optimized independently while maintaining overall system accuracy.
Solution Approach 2:
The patent introduces bottleneck features as an intermediary representation between the spectral features and the Gaussian mixture model classification. These bottleneck features, generated through trained neural networks, condense the spectral information into a compact form that captures essential voice characteristics while reducing the complexity of the final classification task.
Data Source
AI summary
A likelihood of voice that is a discrimination measure between noise and voice is corrected, by using a Gaussian mixture model of noise learned in a time section in which an input signal is noise, and a voice activity is detected on the basis of the corrected likelihood of voice.