We present BARS (Beat-Adaptive Rap Synthesis), a non-autoregressive rapping voice generation model based on Flow Matching, conditioned on a beat track (instrumental backing), lyrics, and a short reference voice for zero-shot timbre control. Unlike prior autoregressive rap synthesis, where rap pace and output duration are strictly determined by the beat, BARS enables independent control over the rap tempo, allowing flexible-length generation without explicit duration prediction or phoneme-level alignment. We encode lyrics using a pretrained ByT5 encoder with a lightweight Transformer adapter trained jointly with the model, without relying on speech-pretrained text encoders, enabling natural handling of non-standard orthography (e.g., slang, abbreviations) and numerals common in rap lyrics. To address alignment instabilities, such as token skipping or repetition, under strong beat conditioning, we propose Asymmetric Scaled RoPE for cross-attention, which introduces a positional structure to facilitate sequential lyric-audio alignment learning. Furthermore, the model supports partial re-generation via span masking, enabling lyric-level editing while preserving the surrounding vocal style. Experiments demonstrate that the proposed model achieves high lyric fidelity, natural rap flow that is rhythmically coherent with the beat, and robust tempo control across a wide range of rapping styles and speeds.

In response to reviewer feedback, we revised the paper as follows.

Abstract/Introduction:

• Clarified the model outputs the vocal track only, not a beat/vocal mix (meta-reviewer)
• defined "beat track" as the instrumental accompaniment (R1).
• Scoped the autoregressive-limitation claim to prior rap-synthesis systems rather than autoregressive modeling in general (R1).

Method:

• Distinguished the paper's core contributions from prior components it builds on (R1).
• Introduced the "A-RoPE" abbreviation, used consistently across text, tables, and figures (meta-reviewer).
• Added a citation to LARoPE, a closely related length-normalized RoPE formulation, and a paragraph contextualizing A-RoPE against it (R1).
• Clarified the adapter's and A-RoPE's distinct, complementary roles.
• Added an intuitive analogy for A-RoPE's position scaling (R2).

Experiment: Implementation Details:

• Added training hardware (8 NVIDIA H100 GPUs) (meta-reviewer)
• Added inference and training details: the Euler ODE solver, bidirectional self-attention, and a 60-second training/inference length cap (R1).
• Added dataset details: artist gender balance, subgenre diversity, and VAE train (R1, R2, R3).

Evaluation Metrics:

• Reframed WER relative to the ground-truth transcription error rate and DNSMOS as a supplementary proxy alongside the subjective MOS study (R3).

Ablation Study:

• Added an observation that a lightweight adapter alone can substitute for costly speech-pretraining.

Training Dynamics:

• Noted A-RoPE's benefit, though not visible in training loss, appears in tempo-mismatch generalization (R2).

Speaker Similarity:

• Clarified that matched/mismatched conditions refer to the same test-set construction as the tempo-robustness analysis (R2),
• Condensed the results table.

Conclusion:

• Clarified the demo page's multilingual examples are preliminary, distinct from the systematic evaluation left to future work (R3).

New section:

• Added an AI Usage Statement (R1, R3).