The use of speech recognition has become increasingly popular in recent years due to its many applications in the realm of human-computer interaction. One of the key challenges in speech recognition systems is the efficient extraction of discriminative features from raw audio signals which directly impacts their performance. This study recommends a distinct method to enhance the precision and resilience of speech recognition systems in which audio feature extraction can be done by VGGish model. This model originally designed for visual recognition tasks, has demonstrated remarkable capabilities in extracting hierarchical and representative features from audio signals. By leveraging the pre-trained VGGish model, we aim to exploit its ability to identify high-level acoustic patterns and transform audio inputs into a more informative feature representation. The proposed method involves two main stages feature extraction and speech recognition. The feature extraction stage involves the use of the VGGish model to process the raw audio signals & creating a feature representation that is both concise and comprehensive. The resulting feature vectors capture various acoustic attributes such as pitch, tempo and spectral content. The mentioned features are subsequently entered into the speech recognition mechanism. To determine the effectiveness of our methodology, we conduct comprehensive examinations on speech recognition datasets that meet industry standards. Our findings suggest that the use of VGGish-based feature extraction significantly enhances the efficiency of speech recognition systems, resulting in greater accuracy and improved resilience to high-noise environments. Furthermore, we conduct a comparative analysis with traditional feature extraction techniques commonly used in speech recognition, such as MFCC and Mel spectrograms