diodiogod-TTS-Audio-Suite/docs/VOCAL_REMOVAL_GUIDE.md

🤐 Vocal/Noise Removal Guide

Complete guide to using the Vocal/Noise Removal node for audio separation and processing.

Overview

The 🤐 Noise or Vocal Removal node is a powerful audio processing tool that can:

• Separate vocals from instrumentals (karaoke creation)
• Remove noise from audio recordings
• Remove echo and reverb artifacts
• Isolate specific audio stems (drums, bass, etc.)

Model Categories & Recommendations

🏆 BEST MODELS (2024-2025)

Top Tier - State of the Art:

• ★★★ MELBAND/denoise_mel_band_roformer_sdr_27.99.ckpt - BEST DENOISING (27.99 dB SDR!)
• ★★★ MELBAND/MelBandRoformer_fp32.safetensors - Kijai's high-quality vocal separation
• ★★★ model_bs_roformer_ep_317_sdr_12.9755.ckpt - Reliable transformer-based (12.97 dB SDR)
• ★★☆ MDX23C-8KFFT-InstVoc_HQ.ckpt - Minimal artifacts (Karafan architecture)
• ★★☆ UVR-MDX-NET-vocal_FT.onnx - Professional vocal extraction (MDX architecture)

📂 VR MODELS (Convolutional Neural Networks)

• UVR-DeEcho-DeReverb.pth - Post-processing: removes echo/reverb artifacts
• HP5-vocals+instrumentals.pth - Balanced vocal/instrumental separation
• 5_HP-Karaoke-UVR.pth - Optimized for karaoke creation (aggressive vocal removal)
• 6_HP-Karaoke-UVR.pth - Alternative karaoke model (different tuning)
• model_bs_roformer_ep_317_sdr_12.9755.ckpt - ⭐ TRANSFORMER ARCHITECTURE (BEST)
• UVR-BVE-4B_SN-44100-1.pth - 4-band processing with spectral normalization
• UVR-DeNoise.pth - Noise reduction specialist

📂 MDX MODELS (Multi-Dimensional eXtraction)

• UVR-MDX-NET-vocal_FT.onnx - Fine-tuned vocal extraction, full-band processing

📂 KARAFAN MODELS (Advanced Hybrid)

• MDX23C-8KFFT-InstVoc_HQ.ckpt - ⭐ HIGHEST QUALITY (8K FFT, minimal artifacts)

🎯 MELBAND ROFORMER MODELS (State-of-the-Art 2025)

Kijai's Models (Vocal Separation):

• MELBAND/MelBandRoformer_fp16.safetensors - Fast processing (456MB, FP16 precision)
• MELBAND/MelBandRoformer_fp32.safetensors - Maximum quality (913MB, FP32 precision)

ZFTurbo's Models (Denoising Specialists):

• MELBAND/denoise_mel_band_roformer_sdr_27.99.ckpt - ⭐ BEST DENOISING (27.99 dB SDR!)
• MELBAND/denoise_mel_band_roformer_aggressive_sdr_27.97.ckpt - Aggressive denoising

Note: MelBand models use advanced transformer architecture with rotary embeddings for superior separation quality

Use Case Workflows

🎤 Karaoke Creation (Vocal Removal)

1. Primary: model_bs_roformer_ep_317_sdr_12.9755.ckpt (aggressive: 12-15)
2. Post-process: UVR-DeEcho-DeReverb.pth (remove artifacts)
3. Use: "remaining" output (clean instrumentals)

🎵 Clean Vocal Extraction

1. Option A: UVR-MDX-NET-vocal_FT.onnx (balanced: 8-12)
2. Option B: MDX23C-8KFFT-InstVoc_HQ.ckpt (highest quality)
3. Use: "extracted voice/noise/echo" output (isolated vocals)

🔧 Audio Denoising

1. Best Quality: MELBAND/denoise_mel_band_roformer_sdr_27.99.ckpt (no aggressiveness setting)
2. Alternative: UVR-DeNoise.pth (gentle: 5-8)
3. Use: "remaining" output (clean audio)
4. Note: "extracted" contains removed noise

💼 Professional Workflow

1. Separate: model_bs_roformer_ep_317_sdr_12.9755.ckpt
2. Clean vocals: UVR-DeEcho-DeReverb.pth
3. Denoise: UVR-DeNoise.pth
4. Combine: Use Merge Audio node

🏠 Beginner-Friendly

• Start with: HP5-vocals+instrumentals.pth (moderate: 8-10)
• Simple setup: No post-processing needed

Model Architectures Explained

⚡ ARCHITECTURE DIFFERENCES

• VR (Vocal Remover): Fast, magnitude-only processing, good for basic separation
• MDX (Multi-Dimensional eXtraction): Hybrid spectrogram/waveform, better quality preservation
• Karafan: Multi-stage ensemble, state-of-the-art quality (highest CPU usage)
• RoFormer: Transformer with rotary embeddings, excellent quality
• MelBand RoFormer: Advanced transformer with Mel-frequency bands, current SOTA (27+ dB SDR)

Aggressiveness Settings (0-20)

📊 RECOMMENDED VALUES

• 0-5: Gentle separation, preserves original audio quality
• 6-10: ⭐ BALANCED (Default: 10) - Good separation with minimal artifacts
• 11-15: Aggressive separation, may introduce artifacts but better isolation
• 16-20: Maximum aggression, highest separation but potential quality loss

🎯 USE CASES BY AGGRESSIVENESS

• 🎤 Karaoke Creation: 12-15 (more aggressive)
• 🎵 Vocal Extraction: 8-12 (balanced)
• 🎼 Preserve Music Quality: 5-8 (gentle)
• 🔧 Problem Audio: 15-20 (maximum effort)

Special Model Types

⚠️ MODELS WITH DIFFERENT BEHAVIOR

UVR-DeNoise.pth - NOISE REDUCTION (not vocal separation)

• "extracted voice/noise/echo" output = extracted noise/artifacts ❌
• "remaining" output = cleaned, denoised audio ✅
• Use case: Remove background noise, hiss, static

UVR-DeEcho-DeReverb.pth - ECHO/REVERB REMOVAL

• "extracted voice/noise/echo" output = extracted echo/reverb ❌
• "remaining" output = dry, processed audio ✅
• Use case: Clean up live recordings, remove room acoustics

HP-Karaoke Models - INVERTED OUTPUTS

• 5_HP-Karaoke-UVR.pth, 6_HP-Karaoke-UVR.pth
• Automatically corrected: System detects and swaps outputs
• Optimized for: Aggressive vocal removal for karaoke

Output Format Options

🎵 AUDIO FORMAT SELECTION

Quality Ranking:

• 📀 FLAC: ⭐ BEST - Lossless compression, perfect quality, larger files
• 🎵 WAV: Uncompressed, perfect quality, largest files
• 🎧 MP3: Lossy compression, smaller files, slight quality loss

Professional Use: FLAC (default) | Fast Workflow: MP3 | Maximum Quality: WAV

File Size Comparison (typical 4-minute song):

• WAV: ~40MB per stem | FLAC: ~20MB per stem | MP3: ~4MB per stem

Advanced Techniques

🚀 MODEL NOTES

• MelBand RoFormer - ✅ NOW SUPPORTED! Best-in-class denoising (27+ dB SDR)
• SCNet models - ⚠️ Architecture issues causing audio artifacts
• MDX23C models - ⚠️ Tensor alignment errors in current implementation
• Note: Aggressiveness setting only affects VR architecture models, not MelBand/MDX/Karafan

💡 PRO TIPS

• Use ensemble combinations for best results
• Apply post-processing models (DeEcho, DeNoise) after primary separation
• RoFormer models are current state-of-the-art (2024-2025)
• Higher quality models require more processing time and memory
• For noise reduction models, use the "remaining" output for your clean audio
• Chain multiple models for professional results

🔄 CACHING SYSTEM

• ✅ ON (Recommended): Saves results to disk, dramatically speeds up repeated processing
• ❌ OFF: Always processes from scratch, ensures fresh results
• Cache includes: Model, aggression, format, and audio content in hash
• Auto-invalidates: When any parameter changes

Troubleshooting

Common Issues

• Empty/silent output: Check if using correct output (remaining vs extracted)
• Poor separation: Try higher aggressiveness or different model
• Artifacts: Use post-processing models or lower aggressiveness
• Slow processing: Enable caching, ensure GPU acceleration

GPU Acceleration

• Required: pip install onnxruntime-gpu (instead of onnxruntime)
• Benefits: 3-10x faster processing for ONNX models
• Check: Look for "ONNX Runtime CUDA acceleration enabled" message

---

For more information, see the main README and Project Index.