Forge Studio 2: How It Works, and Its Limits
Forge Studio 2 is the engine behind every Clean result on AudioForges. This page says what it is, what it was measured against, what changed in the September 2026 update, and where it still falls short, so you can judge it on evidence rather than adjectives.
What it is
A band-split transformer separation model. It splits the audio into frequency bands, reads each band over time with attention, and predicts which energy belongs to the voice, the drums, the bass and the rest. It runs on a dedicated GPU, one song at a time, which is why a run takes about a minute and costs one song. The free result uses a lighter model that finishes in about half a minute and is fine for a rough instrumental or a quick check.
What it was measured against
The comparison page runs one 50-second clip, used with the artist's permission, through the free result, Forge Studio 2, Ultimate Vocal Remover (MDX23C), VocalRemover.org and LALAL.AI. Every output is playable there, every model is named, and the processing times were timed on the same day. On that clip, Forge Studio 2 leaves less audible vocal in the instrumental than the other four. That is one clip; your song may land differently, which is why the free preview exists.
The September 2026 update
- Newer model of the same class, trained on more full mixes.
- Less vocal bleed on backing vocals and reverb tails, the two places the previous model was weakest.
- Cleaner drums and bass in the 4 and 6 stem splits; hi-hats stop leaking into the other stem.
- Same run time and the same one-song price. Every demo clip on the site was re-rendered with it on 28 September 2026.
Where it still falls short
- Dense mixes
- Walls of guitars with the vocal buried in them can leave a faint trace of the voice in the instrumental.
- Live and phone recordings
- Room noise and crowd sound are not a stem the model knows, so they spread across all outputs.
- Heavy reverb
- Very long tails on the vocal can stay with the music. Forge Clean on the vocal stem helps; the instrumental is harder.
- Lyrics and chords
- The engine separates sound. It does not transcribe words or chords. Forge Roll and Forge Score turn a stem into notes, which is a different job.
How to check it yourself in two minutes
- Drop your song on the Vocal Remover and run it free.
- Press the Studio preview. Play, and switch between Free and Studio while it runs; the position does not move.
- Listen to the quietest section, where leftover vocal is easiest to hear. If Studio is clean there, it is clean.
Questions people ask
- Is Forge Studio 2 the same model as LALAL.AI or Moises?
- No. It is our own deployment of a band-split transformer, updated in September 2026, run on a dedicated GPU per song. We publish the same clip through LALAL.AI, VocalRemover.org and UVR so you can compare rather than trust a label.
- Does it remove all vocal bleed?
- On most studio recordings, the instrumental has no audible vocal left. On very dense mixes, live recordings and songs drowned in reverb, a trace can remain. Run it free once and listen to the quiet sections before you decide.
- Why does Clean cost a song when the free result is free?
- The free result runs a lighter model and finishes in about half a minute. Forge Studio 2 runs a heavier model on a dedicated GPU for about a minute per song, and that time is what the song pays for.
- What does the September 2026 update change?
- A newer model of the same class, trained on more mixes. Less bleed on backing vocals and reverb tails, cleaner drums and bass in the 4 and 6 stem splits, and the same run time as before.
- Can I hear it on my own song before paying?
- Yes. Every free run offers a 30-second Clean preview of your own track, and the A/B strip switches between the free result and Clean while it plays.