Newsletter
Join the Community
Subscribe to our newsletter for the latest news and updates
AI platform for synchronized lyrics, karaoke creation, and subtitle conversion
LrcSong is a free web toolkit for AI-powered LRC generation, subtitle conversion, and audio metadata editing. It enables users to strip vocals, sync lyrics, create karaoke videos, and score their singing takes. The platform supports multiple subtitle formats including LRC, SRT, VTT, ASS, and TTML, and offers word-level karaoke synchronization. It is designed to be privacy-first, with client-side tools that never send files to servers.
LrcSong offers a free tier with 3 free credits on signup, enough for roughly 3 AI-generated LRC files. Subtitle converters, viewers, and validators are free and run client-side with no credit cost. Paid plans include Starter, Premium, and Elite, with prices ranging from $0 to $9.99. Elite plan subscribers get a commercial license for all generated outputs, while Starter and Premium plans are for personal use.
Is LrcSong free? Yes. You get 3 free credits on signup — enough for roughly 3 AI-generated LRC files. Subtitle converters, viewers, and validators run client-side and are free with no credit cost.
Are my files stored on your servers? No. Uploaded audio is processed and then automatically deleted, and we never use uploads to train AI models. Client-side tools (converters, viewers, validators) never send files to our servers at all.
How accurate is AI LRC generation? Honest answer: we have a measured number for one engine in our stack, and no number for the one most jobs actually use — so we won't give you a single headline percentage. What we do know: we benchmarked our self-hosted fallback transcriber on Jam-ALT (the peer-reviewed ISMIR 2024 sung-lyrics benchmark) and measured a 60% word error rate at its default settings across 20 songs in English, German, Spanish and French. That's our own worst configuration, and we're publishing it because it's the one that ships. Worth understanding what that number means: word error rate isn't simply 100% minus accuracy. That same run got 63% of the sung words right — the error rate is pushed up mostly by the model adding words nobody sang, which is the typical failure here. Most jobs don't run on that engine; they use a cloud model we haven't benchmarked yet, so quoting any number as your result would be exactly as unsourced as the 90-99% figures advertised across this category, none of which name a dataset or a metric. For scale, on that same benchmark a human transcriber scores 11.1% and the best commercial system 16.1% — sung lyrics are genuinely hard, and anyone claiming 99% is selling you something. One thing that measurably helps: pick your song's language instead of leaving the selector on Auto. Auto guesses wrong on about 1 in 4 songs in our tests, and fixing that was worth ~11 points of accuracy — nearly three times more than vocal isolation. Expect to fix words either way, especially on live recordings or vocals mixed low. Every output is fully editable, and fixing a few words still beats syncing from scratch.
Can I use LrcSong commercially? Yes. Elite plan subscribers get a commercial license for all generated outputs. Starter and Premium plans are for personal use.
What formats do you support? LRC (standard + enhanced), SRT, VTT, ASS/SSA, TTML, and JSON for interchange. Audio input supports MP3, FLAC, WAV, OGG, M4A, and YouTube URLs.
