Feature extraction algorithm optimization in speech recognition

Based on the concept of AI-driven feature modeling, we propose a joint optimization model for multi-scale time-frequency fusion and dynamic feature enhancement tailored for speech recognition tasks. This model utilizes traditional FBank features as its input foundation. By incorporating a multi-scale convolutional architecture, it enables feature modeling of speech signals across varying temporal granularities, simultaneously capturing local phoneme variations and global semantic correlations to enhance the integrity and hierarchical structure of time-frequency features. Furthermore, the model incorporates a temporal attention mechanism to dynamically weight key frames within speech sequences, amplifying responses in salient regions while effectively suppressing silent intervals and noise interference. To balance accuracy and efficiency, the feature extraction module employs deep separable convolutions and channel pruning strategies, significantly reducing computational complexity and parameter count. This enhances the model's deployability in embedded and real-time recognition scenarios. Experiments conducted on the AISHELL-1 and LibriSpeech datasets demonstrate that this model achieves an average word error rate reduction of 1.73% compared to the MFCC+TDNN baseline model, with inference time shortened by approximately 27%. It maintains high recognition stability even under low signal-to-noise ratio conditions. The results demonstrate that multiscale fusion and dynamic enhancement mechanisms exhibit significant synergistic effects in feature extraction, providing technical support for constructing efficient, robust, and portable speech recognition systems.