Hypernasality and decreased clarity are common in people with cleft palate (CP), a birth defect that notably affects speech production. Accurate evaluation of intensity and efficient speech improvement are essential for directing operative treatment and treatment design. Nevertheless, existing methods depend mainly on clinical decisions or limited signal processing, often limited by small datasets and individual differences. This work presents an integrated deep learning system that simultaneously performs the classification of intensity and improvement of speech for CP speech. Manually designed speech features (MFCCs, formants) are combined with deep contextual embeddings derived from a pretrained Wav2Vec2 model. A convolutional neural network (SpectroCNNWithExtras) that was custom-designed enables multi-task learning for both classification of severity and nasality prediction, while a GAN-based SpeechBrain SEGAN system improves spectral quality and reduces hypernasality. Evaluations carried out on the LibriSpeech and NMCPC datasets show notable gains in accuracy of classification, performance of regression, and auditory quality compared to classical signal-based and direct learning methods. The proposed method efficiently reduces limited data availability and inter-speaker differences, delivering a clinically relevant system for dependable evaluation and individualized treatment strategy for patients with cleft palate.