Quantum-Inspired Optimal Fractional-Order Spectrogram for Speech Recognition Enhancement

To address the issues of insufficient resolution and weak feature discriminability in traditional spectrograms for time-frequency representation of speech signals, this paper proposes an optimal fractional-order spectrogram generation method based on a quantum-inspired Newton-Raphson algorithm, aiming to enhance the performance of speech recognition and classification tasks. First, quantum encoding is employed to initialize the population in the Newton-Raphson algorithm. Quantum rotation gates guide individuals to converge toward the optimal solution, while quantum mutation and catastrophe mechanisms are introduced to maintain population diversity and prevent premature convergence. Furthermore, simulated annealing and Lévy flight strategies are incorporated to enhances the algorithm’s global search capability. Next, the audio signal is preprocessed by windowing and framing. A fractional-order spectrogram is generated via the fractional Fourier transform, which is then compressed in scale using a Mel filter bank to obtain a fractional-order Mel spectrogram. Finally, an adjustable fractional-order parameter α is introduced to expand the degrees of freedom in time-frequency representation. By minimizing information entropy as the objective function, hyperparameters including α, frame length, and frame shift are adaptively optimized using the quantum-inspired Newton-Raphson algorithm to obtain the optimal fractional-order spectrogram. Simulation experiment results on the CEC 2022 benchmark functions, the public emotional speech recognition dataset RAVDESS, the sound classification dataset UrbanSound8K, and a self-constructed singing vowel pronunciation dataset demonstrate that, the proposed quantum-inspired Newton-Raphson algorithm exhibits superior global optimization capability and higher stability in solving high-dimensional complex problems compared with existing optimization algorithms. The generated optimal fractional-order spectrograms effectively concentrate signal energy and enhance feature separability, significantly outperforming traditional speech feature extraction methods in terms of accuracy, recall, and F1 score for speech recognition. This method provides a novel approach for high-precision feature extraction of complex speech signals, effectively improves speech recognition performance, and offers good robustness and application prospects.