Making the Most of Limited Data: Score-Aware Training for Text-to-Music Generation
Researchers have developed a novel score-aware training method to improve text-to-music generation, particularly when working with limited data. This technique leverages audio-caption alignment scores as a direct supervision signal, repurposing lower-scoring segments for training. The system, named FluxAudio, also incorporates segment-level filtering and a two-stage captioning process to enhance performance. Submitted to the ICME 2026 ATTM Grand Challenge, the 450M-parameter model achieved strong results, ranking second in objective evaluation and third in the efficiency track. AI
IMPACT This score-aware training method could enable more efficient development of text-to-music models, reducing reliance on massive datasets.