An effective cluster-based model for robust speech detection and speech recognition in noisy environments.

Area of Science:

• Signal Processing
• Speech Technology
• Machine Learning

Background:

• Speech recognition systems struggle in noisy environments.
• Accurate voice activity detection (VAD) is crucial for improving speech recognition performance.
• Existing VAD methods have limitations in characterizing noisy channels.

Purpose of the Study:

• To develop an accurate speech detection algorithm for noisy environments.
• To improve the performance of automated speech recognition (ASR) systems.
• To reduce computational cost for real-time applications.

Main Methods:

• A hard decision clustering approach with prototypes to model noisy channels.
• A decision rule based on averaged distance to a cluster-based noise model.
• Utilizing contextual information (speech frame neighborhoods) for robust detection.

Main Results:

• Significant improvements in detection accuracy compared to standard VADs.
• Enhanced speech recognition rates demonstrated on AURORA 2 and AURORA 3 databases.
• Reduced computational cost, making the algorithm suitable for real-time ASR.

Conclusions:

• The proposed speech detection algorithm effectively handles noisy environments.
• Contextual information and clustering enhance VAD robustness and accuracy.
• The algorithm offers a computationally efficient solution for real-time speech recognition.