← All topics View on GitHub ↗

💬 General

Karaoke Maker — building the file when the song has none

StartSWest ·

This thread is for the Karaoke Maker.

What it's for: making the karaoke file when the song does not have one — from the song and nothing else.

How it works: it can first split the song into two tracks on your own machine: the backing track to sing over, and the voice, which is what the transcriber and the melody reader listen to. It takes well under a minute with a usable graphics card, about six without. The words can be transcribed from the audio or pasted in and kept. Timing is recorded by ear — mark each line as the track plays, then individual words where a line needs the detail. The melody is read off the recording into notes you can hear, split, retune, delete or add to by hand. Undo and redo run throughout, and the result goes straight to the player or saves as a project.

What it downloads, and only when you ask: a speech model to transcribe (about 570 MB, or 1.1 GB without a capable graphics card), a 361 MB pitch model to read the melody, and a 700 MB model to split the song. Each comes down once, on your press, never at launch, and your audio never leaves the machine. Pasting the words in, bringing your own instrumental, or timing by ear skips them.

Ask here about splitting, transcription, timing, the melody editor and project files.


Rewritten for 2.1.

Formatting: **bold** _italic_ `code` - list [link](url)

Images: the box here takes text only — giscus cannot upload files. To attach a screenshot, write the post on GitHub and it appears here.

Loading the discussion from GitHub…

Topics Read the FAQ first

New to FluidEQ?

FluidEQ in motion