Dysphonia is a condition characterized by difficulties in voice production caused by functional, psychological, and neurological factors. Accurate diagnosis of dysphonia is crucial for determining the proper treatment procedures and follow-ups. The previous approaches primarily focused on sustained vowels and manual feature extraction, which does not represent everyday speech usage. It raised questions regarding the usability of these methodologies when applied in real-life scenarios. Also, differentiating between organic and functional dysphonia remains an unexplored area in speech pathology. In this paper, we propose an approach based on Web-scale Supervised Pretraining for Speech Recognition (Whisper). Features extracted from the pre-trained transformer-based Whisper encoder in both Base and Large variations were used to train machine learning models such as Support Vector Machine, Random Forest, and Multi-Layer Perceptron. We also proposed an architecture based on a 1-dimensional convolution neural network (1DCNN). Our proposed method showed high performance and surpassed previous approaches on the same dataset in binary and multiclass classification. It achieved 95.51% accuracy in binary classification and 76.40% in multiclass classification, outperforming previous methods. These results emphasize the effectiveness of our model in capturing speech characteristics at the utterance level and distinguishing between dysphonia subtypes.