This work presents MambaVoc, a high-fidelity audio synthesis model built upon state-space model (SSM) architectures. MambaVoc adopts a bidirectional variant of Mamba by replacing the internal causal convolution with non-causal 1D convolution, enabling joint modeling of both past and future audio contexts. Compared to the original unidirectional Mamba, the bidirectional design captures richer temporal dependencies, while 1D convolution enhances local feature extraction without causal constraints. We evaluate MambaVoc’s robustness under unseen conditions using VCTK, AISHELL-3, and MUSDB18-HQ datasets. The results demonstrate that the model maintains high audio quality while generalizing well across various domains. Ablation studies confirm the complementary effects of the bidirectional SSM and 1D convolution in modeling both global spectral structures and fine-grained waveform details. Additionally, MambaVoc exhibits zero-shot adaptation to unseen speakers and audio domains, underscoring its potential for real-world deployment.