next-generation lyric sync
automatic lyric syncing that just works. not only to the line, not even to the word, but to the syllable.
unmodified, 100% automatic sync
accuracy
autolrc is state of the art and still improving. not only does it have a powerful base, it also introduces a novel technique: line anchors. a lot of songs already have line level lrc files, so we can use it as anchors to enhance with our deeper level sync. for JamendoLyrics (standard eval):
| method | median error | mean error | words ≤300ms off ↓ |
|---|---|---|---|
| autolrc + line anchors | 32 ms | 70 ms | 97.2% |
| autolrc | 34 ms | 190 ms | 95.8% |
| dafx (2025) | 41 ms | 216 ms | 95.2% |
| kglw (2025) | — | 200 ms | 94% |
| gupta et al. (2020) | 50 ms | 220 ms | 94% |
| huang et al. (2022) | — | 230 ms | 94% |
| hx-ih (2023) | 43 ms | 160 ms | 93% |
| mstre-net (2021) | — | 310 ms | 93% |
| spotify (2023) | — | 150 ms | 92% |
| vaglio et al. (2020) | — | 370 ms | 92% |
| hx-d (2023) | 43 ms | 420 ms | 89% |
| stoller et al. (2019) | — | 820 ms | 85% |
faq
what?
autolrc creates time synced lyrics to the syllable for any given song. lyric sync is almost always manually done today, and this makes it easy and possible in under a minute.
how can I use it?
right now, you cannot. it is still being actively developed and entirely as a fun, few hours a month project. at some point it will be released fully open-source alongside a paper. if you know me dm for possible access :)
what does it need?
just an audio file and lyrics text or the song name (it can automatically search for and download the lyrics if available). autolrc does not speech recognition as it is much less accurate, so requires the lyrics.
does it work on rap? metal? etc
it can work surprisingly well, even with difficult vocals. it depends entirely on the situation, it is impossible to know how well autolrc will do just from listening.
what about languages other than english?
only english is officially supported/tested but in theory with some effort this can work for most languages.
how long does it take to sync a song?
if there are line anchors, autolrc can stream to the client. this makes the first fully synced line appear in under 1 second and then stream at ~6x realtime. in total, it takes 20-30s for typical song lengths. without anchors, there is no stream and it typically takes 40-80s.
proof for benchmark results?
when autolrc is publicly released, it will be completely open source and be published with a paper showing benchmark details and methodology.
why syllable-level instead of word?
autolrc technically has phoneme-level but it is less reliable. syllables are not only nicer to watch animate than entire words but unlock some fun features for music clients and the future.