Why vocals first
Speech models are trained mostly on talking. In a full mix, drums, guitars and reverb cover consonants and the model guesses or skips lines. With the vocal stem alone the words stand out, which is why the order is: split, then transcribe.
| Input | Typical result |
|---|---|
| Vocal stem, clear singing | Most lines right; fix a few words |
| Full mix, quiet backing | Usable draft, more errors |
| Full mix, loud band, screams or rap at speed | Many wrong or missing lines |
| Heavy harmonies or choirs | Words merged; edit by ear |
Honest limits
- Whisper can mis-hear lyrics, invent words in instrumental parts and repeat lines. Always listen along.
- It does not look lyrics up: it only writes what it hears from your file.
- Lyrics of released songs are copyrighted. Use this on your own songs, covers you have the rights to, or for personal study; do not publish other people's lyrics.
- 1 free transcription a day, then 5 credits each.
We do not store your audio. Your browser shrinks it to a small file, sends it over HTTPS to our API, and we pass it to our AI provider (fal.ai) for the job; the result comes back to your browser. Do not use it for recordings you are not allowed to share with a processor.
What to do with the lyrics
- Download SRT and use it in the lyric video maker or your video editor.
- Edit the lines and send them to Lyrics to song to hear them sung in a new style.
- Keep the timestamps to sync karaoke or captions.
Frequently asked questions
Can it get lyrics from any song?
It writes what it hears in the file you give it, with errors. It does not look up released lyrics, and those lyrics are copyrighted, so do not republish them.
Should I remove the music first?
Yes, it is the single biggest improvement. Use the vocal isolator, then transcribe the vocals file.
Does it work for rap and fast singing?
It can, but fast or mumbled delivery causes more mistakes. Expect to edit.
Which languages are supported?
About 100, including English, Spanish, French, German, Portuguese, Japanese, Korean and Chinese. Set the language for singing, since automatic detection can be fooled by music.
Can I get synced timestamps?
Yes, one per line. They come from the speech model and may drift by a second or so; SRT and VTT download are included.
How much does it cost?
The first run each day is free; after that 5 credits.
Is my song stored?
Not by us. It is processed by our AI provider, fal.ai, for the job.