Related Experiment Videos
Hierarchical singleton-type recurrent neural fuzzy networks for noisy speech recognition.
Chia-Feng Juang1, Chyi-Tian Chiou, Chun-Lung Lai
1Department of Electrical Engineering, National Chung-Hsing University, Taichung 402, Taiwan. cfjuang@dragon.nchu.edu.tw
IEEE Transactions on Neural Networks
|May 29, 2007
Summary
This study introduces a novel hierarchical recurrent neural fuzzy network (HSRNFN) for improved noisy speech recognition. This method effectively filters noise and recognizes speech patterns, outperforming traditional models.
Area of Science:
- Artificial Intelligence
- Signal Processing
- Machine Learning
Background:
- Speech recognition systems struggle with performance degradation in noisy environments.
- Existing methods like MLPs, TDNNs, and HMMs have limitations in handling complex noise patterns.
- Recurrent neural fuzzy networks offer potential for temporal pattern processing in speech.
Purpose of the Study:
- To propose a novel Hierarchical Singleton-type Recurrent Neural Fuzzy Network (HSRNFN) for robust noisy speech recognition.
- To enhance speech recognition accuracy by integrating noise filtering and recognition capabilities within a unified framework.
- To evaluate the effectiveness of the HSRNFN against established speech recognition models.
Main Methods:
- Developed an HSRNFN by hierarchically connecting two Singleton-type Recurrent Neural Fuzzy Networks (SRNFNs).
- One SRNFN is dedicated to noise filtering, while the other handles word recognition.
- SRNFNs utilize recurrent fuzzy if-then rules with fuzzy singletons for temporal speech pattern processing.
- Implemented a recognition criterion based on the prediction error of individual SRNFN word models.
Main Results:
- The proposed HSRNFN demonstrated superior performance in Mandarin word recognition tasks under various noise conditions.
- Experimental comparisons showed that HSRNFN outperformed Multilayer Perceptrons (MLP), Time-Delay Neural Networks (TDNNs), and Hidden Markov Models (HMMs).
- The integrated noise filtering mechanism significantly improved the robustness of the speech recognition system.
Conclusions:
- HSRNFN provides an effective approach for noisy speech recognition, offering improved accuracy and robustness.
- The hierarchical structure effectively separates noise filtering and recognition tasks, enhancing overall system performance.
- This novel architecture presents a promising advancement for real-world speech recognition applications facing environmental noise.