A recurrent neural fuzzy network for word boundary detection in variable noise-level environments

Area of Science:

• Speech Processing
• Signal Analysis
• Artificial Intelligence

Background:

• Automatic word boundary detection is challenged by variable background noise.
• Existing algorithms often fail due to fixed noise level assumptions.
• Noise variability during recording significantly impacts algorithm performance.

Purpose of the Study:

• To develop a robust word boundary detection algorithm for variable background noise.
• To introduce a novel approach overcoming limitations of current methods.
• To enhance speech recognition accuracy in noisy environments.

Main Methods:

• Proposing a refined time-frequency (RTF) parameter for feature extraction from noisy speech.
• Extending the time-frequency (TF) parameter to multiband spectrum analysis for clearer signal-noise distinction.
• Developing a recurrent self-organizing neural fuzzy inference network (RSONFIN) for temporal relation processing and noise variation adaptation.

Main Results:

• The RTF-based RSONFIN algorithm successfully detects word boundaries amidst variable background noise.
• Achieved a higher recognition rate (approx. 12% improvement) compared to the TF-based algorithm in variable noise conditions.
• Reduced endpoint detection error rate to 23%, significantly outperforming the TF-based algorithm's 47% average.

Conclusions:

• The proposed RTF-based RSONFIN algorithm offers superior performance for word boundary detection in variable background noise.
• RSONFIN's self-learning and economic network size contribute to high learning speed and adaptability.
• This method provides a significant advancement over existing algorithms for noisy speech processing.