Speech endpoint detection based on speech time-frequency enhancement and spectral entropy

Area of Science:

• Speech processing
• Signal processing
• Acoustics

Background:

• Accurate endpoint detection is critical for speech recognition systems.
• Non-speech regions in audio input can degrade performance.
• Noisy environments pose significant challenges for reliable endpoint detection.

Purpose of the Study:

• To propose a novel and robust approach for speech endpoint detection.
• To enhance speech signals in noisy environments for improved accuracy.
• To develop a method suitable for real-time digital signal processing (DSP) systems.

Main Methods:

• Integration of time-frequency enhancement and spectral entropy features.
• Application of spectral subtraction in the frequency domain to remove additive noise.
• Utilizing a weight function based on short-time energy and zero-crossing rate in the time domain to mitigate residual noise.
• Employing a spectral entropy-based method for precise endpoint detection by monitoring feature transitions.

Main Results:

• The proposed algorithm demonstrates robustness across various noise types, particularly in low signal-to-noise ratio (SNR) conditions.
• The method effectively enhances speech signals, leading to more precise endpoint detection.
• The algorithm exhibits low computational complexity, making it suitable for real-time applications.

Conclusions:

• The integrated approach of spectral enhancement and spectral entropy provides a robust solution for speech endpoint detection.
• The method is highly effective in noisy environments and low SNR scenarios.
• The algorithm's low complexity ensures its applicability in real-time digital signal processing systems for speech recognition.