Mass spectrometry (MS) is a powerful tool for analyzing small molecules in complex samples, yet accurate structure elucidation from spectral data remains challenging. We present a Transformer-based model for de novo molecular generation, named MoleNovo, which directly predicts SMILES strings from MS spectra. MoleNovo employs a tailored embedding strategy for both mass-to-charge ratio (m/z) and intensity values within a sequence-to-sequence architecture. Evaluated on the NIST20 dataset, MoleNovo outperforms existing models such as Spec2Mol and MS2Mol across multiple metrics, including BLEU score, Levenshtein distance, fingerprint similarity, and chemical validity. These results underscore the potential of Transformer architectures for scalable, automated compound identification in metabolomics and drug discovery.