Abstract
Aiming at the problem of low recognition rate in speech recognition methods, a speech recognition method in multi-layer perceptual network environment is proposed. In the multi-layer perceptual network environment, the speech signal is processed in the filter by using the transfer function of the filter. According to the framing process, the speech signal is windowed and framing processed to remove the silence segment of the speech signal. At the same time, the average energy of the speech signal is calculated and the zero crossing rate is calculated to extract the characteristics of the speech signal. By analyzing the principle of speech signal recognition, the process of speech recognition is designed, and the speech recognition in multi-layer perceptual network environment is realized. The experimental results show that the speech recognition method designed in this paper has good speech recognition performance.
Introduction
Despite the ubiquity of voice-activated technology, speech recognition still faces a significant "noise floor" problem. While machines perform well in quiet environments, real-world "multi-layer perceptual network environments" introduce environmental noise and transmission distortions that degrade signal quality. Traditional models often rely on static features that fail to capture the transient, dynamic nature of speech, and fixed learning rates in neural networks frequently lead to slow convergence or local minima.
This research presents a robust speech recognition framework that combines traditional digital signal processing (DSP)—specifically Mel-frequency cepstrum coefficients (MFCC) and their first-order differences—with a multi-layer perceptual network based on Restricted Boltzmann Machines (RBM). By integrating dynamic speech features and an adaptive learning rate, the proposed method achieves high recognition accuracy on isolated word tasks.
Methodology
Signal Preprocessing and Feature Extraction
The system follows a hybrid pipeline that blends classic DSP with modern deep learning. The process begins with an analog voice signal that undergoes sampling and quantization. To improve the signal-to-noise ratio, a high-pass filter ($H(z) = 1 - aZ^{-1}$) is used to "pre-emphasize" the signal, followed by windowing and framing.
A critical step in this pipeline is endpoint detection. The system utilizes a "double threshold" algorithm based on average energy and zero-crossing rates to strip away silence segments, ensuring the model only processes meaningful audio data. For feature extraction, the authors calculate both standard MFCCs (static power spectrum) and the first-order difference of MFCCs ($\Delta$MFCC). This "mixed" approach captures the velocity of spectral changes, providing a dynamic signature that is crucial for distinguishing similar-sounding phonemes.
Multi-Layer Perceptual Network Architecture
The core of the recognition engine is a network utilizing Restricted Boltzmann Machines (RBM). The architecture consists of an input layer, a hidden layer with specific transfer functions, and an output layer for speech recognition.
Key innovations in the network include:
- RBM Pre-training: RBMs are used to initialize weights, helping the model find a stable state before final classification.
- Adaptive Learning Rate: To prevent overshooting and speed up training, the learning rate is adjusted dynamically. If the error function decreases, the rate increases by 5%; if the error increases, the rate is reduced by 25%.
- Softmax Classification: A Softmax classifier at the top layer maps processed features into specific word categories.
Experimental Results
The method was evaluated using a custom database of 500 samples (5 speakers pronouncing Chinese digits 1–10, sampled at 8kHz). The results demonstrated clear superiority for the hybrid approach:
- Feature Performance: Using the combined MFCC + $\Delta$MFCC features achieved 87.2% accuracy, compared to 82.0% for static MFCCs alone.
- Overall Accuracy: The proposed system correctly identified 196 out of 200 test samples, resulting in a 98% recognition rate.
- Comparative Analysis: The method significantly outperformed contemporary benchmarks, including CNN-based models (82%) and other deep learning references (91%).
Conclusion
This research demonstrates that accounting for the dynamic "motion" of audio signals through $\Delta$MFCC and utilizing an adaptive multi-layer perceptual network can significantly improve speech recognition performance. While the current study focuses on isolated words, future work will aim to transition to continuous speech recognition and explore further optimizations to balance computational complexity with real-time performance.
Summary updated