Speech Classification Audio Wake on Voice Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speech/non-speech detection techniques suffer from low accuracy and high false positive and false negative rates, particularly in noisy environments, which is critical for implementing wake on voice and always listening devices.
Innovation Solution
A computer-implemented method using a deep neural network-based acoustic model that generates probability scores for audio units, updating a speech pattern model with self-loops and transitions to accurately differentiate between speech and non-speech, enabling efficient detection and classification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If sample-based voice activity detection using audio signal characteristics (short term energy, zero crossing rates) is used, then the detection system is simple to implement, but the accuracy is low with high false positive and false negative rates
Solution Approach 1:
The patent replaces traditional signal processing methods (mechanical/system-based analysis of short term energy and zero crossing rates) with a deep neural network-based acoustic model. This substitution enables the system to achieve high accuracy in speech/non-speech detection by learning complex patterns from audio data, while still maintaining practical implementability through software-based deployment.
2Loss of information
If frequency-based voice activity detection with fast Fourier transform is used, then frequency domain analysis is provided, but the accuracy remains low with similar limitations
Solution Approach 1:
The patent transforms the detection approach by changing from traditional frequency domain parameters (obtained via fast Fourier transform) to probabilistic parameters generated by a deep neural network acoustic model. The model outputs probability scores for different audio units (speech, non-speech, noise, silence), enabling more accurate classification while preserving frequency domain analysis capabilities through the neural network's learned features.
3Measurement precision
If deep neural network-based acoustic model with speech pattern model is used, then speech detection accuracy is significantly improved, but the computational complexity increases
Solution Approach 1:
The patent segments the audio signal into discrete audio units and processes each unit through the acoustic model independently. The speech pattern model further segments the detection process into distinct states (speech, non-speech, noise, silence) with defined transitions. This segmentation approach enables parallel processing and reduces the computational burden compared to analyzing the entire audio stream as a single continuous signal.
Solution Approach 2:
The patent uses a speech pattern model that copies and extends the acoustic model's probability outputs into a structured state transition framework. Instead of directly using raw neural network outputs, the system creates a simplified symbolic representation (speech pattern model) that captures the essential detection logic, making the complex neural network results more manageable and interpretable.
4Use of energy by moving object
If traditional voice activity detection is used in noisy environments, then the system operates with low resource requirements, but the false positive and false negative rates increase
Solution Approach 1:
The patent replaces traditional low-resource signal processing methods with a deep neural network-based acoustic model that is specifically trained to handle noisy environments. The model learns to distinguish speech from noise and other interfering sounds by training on diverse audio data including noisy conditions, thereby achieving high reliability while maintaining reasonable resource requirements through efficient neural network deployment.
Data Source
Figure 1~2
Figure 3~4
Figure 5~6
AI summary
Speech or non-speech detection techniques are discussed and include updating a speech pattern model using probability scores from an acoustic model to generate a score for each state of the speech pattern model, such that the speech pattern model includes a first non-speech state having multiple self loops each associated with a non-speech probability score of the probability scores, a plurality of speech states following the first non-speech state, and a second non-speech state following the speech states, and detecting speech based on a comparison of a score of the first non-speech state and a score of the last speech state of the multiple speech states.