Text2midi-InferAlign: Improving Symbolic Music Generation with Inference-Time Alignment

We present Text2midi-InferAlign, a novel technique for improving symbolic music generation at inference time. Our method leverages text-to-audio and music-structural alignment rewards to guide the generation process, encouraging consistency with the input caption. Specifically, we introduce two objective scores: a text-audio consistency score measuring rhythmic and semantic alignment, and a harmonic-consistency score penalizing notes outside the musical key. By optimizing these objectives during generation using a reward-guided tree search, our model produces symbolic music more faithful to the input captions. Our approach can extend any existing autoregressive model without requiring retraining. We evaluate our work on top of the Text2midi model, demonstrating significant improvements in both objective and subjective metrics.