Real-Time Speech Endpoint Detection via Spectral Entropy Analysis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
State-of-the-art speech activity detection systems face variability in performance due to characteristics of non-speech and background noises in audio signals, affecting the accuracy of endpoint detection.
Innovation Solution
A lightweight method and system that detects speech endpoints in real-time by analyzing entropy statistics of audio signal frequency spectra, involving buffering, adding random noise, filtering, and calculating entropy to determine speech and non-speech portions, with onset and offset thresholds to identify speech start and end points.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If state-of-the-art systems utilize features such as energy, zero-crossing, and frequency spectrum to detect speech, then speech detection capability is achieved, but performance varies significantly depending on background noise characteristics
Solution Approach 1:
The patent transforms the audio signal into the frequency domain using Fourier transform and computes spectral flux as a new parameter. This parameter change from time-domain features (energy, zero-crossing) to frequency-domain features (spectral flux, entropy) enables the system to capture noise characteristics more effectively and achieve better adaptability across different acoustic environments while maintaining reliable speech detection
2Measurement precision
If pre-trained models are employed to distinguish speech and non-speech frames, then detection accuracy improves, but computational complexity increases
Solution Approach 1:
The patent extracts specific key features (spectral flux, entropy, zero-crossing rate) from the audio signal and uses these extracted features directly for speech endpoint detection. By taking out only the most discriminative features rather than using entire pre-trained models, the system achieves good detection accuracy while significantly reducing computational complexity and making the system more suitable for real-time processing
Solution Approach 2:
The patent employs lightweight, computationally inexpensive features (spectral flux, entropy calculations) that can be computed quickly and discarded after use, replacing expensive pre-trained models. These simple statistical features provide sufficient discrimination power for speech endpoint detection while being computationally efficient and suitable for resource-constrained environments
3Productivity
If traditional speech activity detection methods are used, then basic speech detection is achieved, but robustness to noise variations deteriorates
Solution Approach 1:
The patent performs preliminary analysis by computing the average spectral flux and entropy over a background period before speech detection begins. This preliminary action establishes a noise baseline that is then used to normalize subsequent speech detection, enabling the system to maintain high robustness to noise variations while keeping the actual detection process fast and efficient
Data Source
AI summary
Systems and methods for detecting speech activity. The system includes an audio source and an electronic processor. The electronic processor is configured to receive a first audio signal from the audio source, buffer the first audio signal, add random noise to the buffered first audio signal, and filter the first audio stream to create a filtered signal. The electronic processor then determines a signal entropy of each frame of the filtered signal, determines an average signal entropy of a first plurality of frames of the filtered signal occurring at a beginning of the filtered signal, and compares the signal entropy of each frame of the filtered signal to the average signal entropy. Based on the comparison, the electronic processor determines a first speech endpoint located in a first frame of the filtered signal.