A Recursive Search Method for Lyrics Alignment
Emir Demirel, Sven Ahlbäck, Simon Dixon · Zenodo (CERN European Organization for Nuclear Research) · 2020
Audio-to-lyrics transcription and alignment requires strong acoustic and language models. Even in the presence of such models, the length of audio segments for decoding remains a challenge. In this year’s MIREX submission, we present a recursive search method that splits the audio with respect to anchoring words for performing alignment on shorter audio segments. The recursion is applied by gradually restricting the language model and search space after each search iteration. We apply a final pass of forced alignment on the segmented audio to obtain timings for every word in the input song lyrics. According to initial experiments, our system is robust to various musical genres while being executable on local machines with low memory and computational resources.