Skip to content

AI Models Guide

11. Guide to AI Models in EarPurifier

EarPurifier integrates three world-class neural networks engineered for vocal isolation and background music removal. Each model is tailored to specific hardware capabilities and quality requirements:

1. RoFormer (The Gold Standard · Highest Fidelity)

  • Separation Quality: Ranked #1 worldwide for isolating human speech from complex music and percussion with unmatched vocal clarity.
  • Best Suited For: Heavy background music, loud beats/drums, documentary archives, and tracks where pristine vocal preservation is paramount.
  • Hardware Profile: Computationally demanding. Recommended for Cloud Mode (Kaggle T4), or locally only if your machine is equipped with a powerful modern GPU.

2. MDX23C (The Balanced Choice · Quality & Speed)

  • Separation Quality: Delivers clean, crisp vocal separation while processing roughly twice as fast as RoFormer.
  • Best Suited For: Standard educational video lectures, interviews, and general documentary narrations with moderate music.
  • Hardware Profile: Balanced. Runs fast and efficiently on Cloud Mode and on mid-tier desktop graphic cards.

3. Kim Vocal 2 (Ultra-Fast · Lightweight)

  • Separation Quality: Clean vocal isolation designed to preserve speech clarity; delivers roughly 80% of RoFormer's top-tier quality at lightning-fast execution speed.
  • Best Suited For: Extended podcasts, multi-hour lectures, and running locally on standard laptops or PCs without a dedicated GPU.
  • Hardware Profile: Extremely lightweight. Runs smoothly on standard computer processors (CPUs) without requiring a graphics card.
  • Cloud Advantage: This is the only model that works in Cloud Mode even if your Kaggle account is not phone-verified, as it can run on Kaggle's free CPU instances!

How to Choose the Right Model? (Decision Tree)

Model Selection Decision Tree

Next: Troubleshooting & FAQ.