We introduce a fully differentiable end-to-end audio transformation network designed to convert the style of one audio sample to another. This method offers three significant advantages: (a) it operates without the need for parallel utterances, transcriptions, or time alignment processes; (b) it utilizes a global conditioning mechanism, making it vocabulary agnostic and capable of transforming audio styles regardless of the target identity; and (c) it performs one-shot audio transformations without intermediate phonetic representations, thus eliminating the necessity for phonetic alignments and speaker-independent ASR networks. We assess our method against existing approaches in voice conversion and musical style transfer tasks. Subjective evaluations demonstrate the superiority of our approach. The network employs an encoder-decoder architecture that integrates neural network models known for their explainability in natural language processing tasks.