Text-to-music generation has advanced but remains constrained by Western harmonic structures, limiting its applicability to non-Western traditions like the heterophonic Nanyin music. This paper introduces Narrative2Nanyin, a novel symbolic music generation system that creates multi-track MIDI compositions for Pipa, Sanxian, Dongxiao, and Erxian directly from narrative text. The system uses a culturally grounded tokenization scheme, NanyinTok, to capture pitch, rhythm, and instrument-specific techniques without relying on chords. Built on a BERT–GPT2 encoder–decoder architecture, the model is fine-tuned via Low-Rank Adaptation (LoRA) for efficient adaptation. An interactive interface enables expert-in-the-loop evaluation. Automatic evaluations demonstrate strong performance, with a Scale Adherence score of 0.803 and Dynamic Range score of 0.793, indicating the model’s ability to generate pentatonic and narratively expressive music. This approach advances culturally inclusive music generation by abstracting heterophony as a reusable representation.