I've noticed that many users recommend a series of modules to achieve a desired stem separation, for example:

UVR-MDX-NET Karaoke + VR5_HP-Karaoke-UVR
Or
MDX-NET HQ 5 + Inst Main
Or
The Attached Chart, which really confuses me.

When I put more than one module in Ensemble mode, I get a bunch of stems . . . like it's stemming all the modules and giving me choices for comparison. Is there a way to have a few modules in a series, i.e., A+B+C=Stems from the combination of A,B,C processed in a row??

Any help and/or thoughts are appreciated.

 

im sorry i don't understand your question correctly and i cant answer it...

but remember, that graph is old... there is new models now... they are good and no need for Ensembling...

my suggestion:
just use becruily inst and vocal
inst stems: online: mvsep-> bs roformer sw / offline : zftorbo bs roformer or htdemucs_ft (you can Ensemble all demucs ones) / the mvsep one is best that you can get for now
reverb: anvuew models
karaoke: becruily

this doc can help you( but dont lost in it):

https://docs.google.com/document/d/17fjNvJzj8ZGSer7c7OFe_CNfUKbAxEh_OBv94ZdRG5c

 

Thanks so much for taking the time to share such a clear, comprehensive explanation. MUCH appreciated! If not too much trouble, would you kindly provide your suggestion for the following two types of separation?

1. Vocals. Whatever cleanly separates them from the entire instrumental track. I don't mind lead and backing sandwiched together . . . just as long as they are both pulled out. Yet, if there is a way to also separate lead and harmonies after the fact . . . that would also be useful.

2. Drums. I'm desperately trying to extract drum loops from full performances. I use these for practice, quick song assembly, and as a teaching tool.

Any thoughts or suggestions would be appreciated. Note: I realize that one model and associated settings will not always yield the best results in all situations. I'm hopeful to find a workflow and parameters that give me the best chance(s).

Finally, in your opinion, is UVR the best for separating at this moment in time? Others seem to package it all in easy to use interfaces, but are any of the alternatives better at doing any kind of separation?

THANK YOU!!!!!!

 

Last edited: Aug 5, 2025

Spoiler: Source: Grok KI

I understand you're looking for guidance on using Ultimate Vocal Remover (UVR) for vocal and drum separation, along with an evaluation of whether UVR is the best tool for these tasks in 2025. Below, I’ll address your questions in detail, providing recommended workflows and settings for vocal and drum separation, and comparing UVR to other available tools based on the latest information. I’ll keep it concise yet comprehensive, focusing on practical advice for your use cases (extracting vocals and drum loops for practice, song assembly, and teaching).

1. Vocal Separation with UVR
Goal: Cleanly separate vocals (lead and backing combined) from the instrumental track, with an optional method to further separate lead and backing vocals.

Recommended Workflow for Vocal Separation
UVR is highly effective for vocal separation due to its advanced AI models, particularly the MDX-Net and VR architectures. Here’s a step-by-step workflow to achieve clean vocal extraction:

1. Model Selection:
   • Primary Model: Use the MDX23C-InstVoc HQ model for high-quality vocal/instrumental separation. This model is praised for its full-band processing (unlike Spleeter’s 11kHz cutoff) and minimal artifacts.
   • Alternative Model: If MDX23C-InstVoc HQ isn’t available in your UVR version, use Kim Vocal 2 (MDX) for clean vocal stems. It’s noted for producing high-quality results with fewer processing steps.
   • Ensemble Mode: For the best results, enable UVR’s ensemble mode, combining MDX23C-InstVoc HQ with Kim Vocal 1 or Voc-FT. This averages outputs from multiple models to reduce artifacts and improve clarity.
2. Settings:
   • Segment Size (Chunks): Set to default (e.g., 32 or 64). Larger segments (e.g., 128) improve quality but require more GPU memory and processing time.
   • Overlap: Use a moderate overlap (e.g., 0.5–0.8) to balance quality and processing speed. Higher overlap (e.g., 0.9) can reduce artifacts but increases computation time.
   • Output Format: Export in WAV for the highest quality, especially for further editing or remixing. UVR supports MP3, FLAC, and WAV, but WAV minimizes quality loss.
   • GPU Acceleration: If you have an NVIDIA GPU with CUDA cores (e.g., RTX 1060 or better), enable GPU rendering for faster processing (e.g., 30 seconds for a 3–4 minute song). CPU rendering is viable but slower.
3. Post-Processing for Cleaner Vocals:
   • Run the vocal output through VR De-Reverb and VR De-Echo-Normal models sequentially to reduce reverb and echo artifacts. This is particularly effective for cleaner vocals without metallic or “blip-shatter” sounds.
   • Example Workflow: Process with MDX Kim Vocal 2, then apply VR De-Reverb, followed by VR De-Echo-Normal. This simplifies the process while maintaining quality.
   • Note: Some users report metallic artifacts with de-echo models. If this occurs, try only VR De-Reverb or adjust input audio quality (higher-quality inputs like 44.1kHz WAV yield better results).
4. Optional: Separating Lead and Backing Vocals:
   • UVR includes a Lead & Back Vocal Splitter option in some models, which can separate lead vocals from backing vocals. Use the MDX-B Karaoke model or check MVSEP’s Medley Vox model for multi-singer separation. These models are designed for high-quality lead vocal extraction and can handle backing vocals to some extent.
   • Process: After extracting the vocal stem (lead + backing), run it through the Lead & Back Vocal Splitter in UVR or MVSEP’s Medley Vox model. Note that this works best with tracks where vocals are distinct (e.g., no heavy reverb blending lead and backing).
   • Limitations: If lead and backing vocals are heavily mixed or similar in frequency, separation quality may degrade. Test with a high-quality input file (e.g., 44.1kHz WAV).

Tips for Success

• Input Quality: Use high-quality audio (e.g., 44.1kHz WAV or FLAC) to minimize artifacts. Lower-quality MP3s may leave faint vocal traces or introduce distortion.
• Experimentation: Vocal separation quality varies by song due to mixing styles. Test multiple models (e.g., MDX23C, Kim Vocal 2, Voc-FT) and ensemble combinations for each track.
• Cleaning Up Artifacts: If artifacts persist, import the vocal stem into a DAW (e.g., Audacity, Logic Pro) and apply light EQ or noise reduction to remove residual instrumental bleed.

2. Drum Separation with UVR
Goal: Extract clean drum loops from full performances for practice, song assembly, and teaching.

Recommended Workflow for Drum Separation
UVR’s Demucs v4 models are particularly effective for drum separation, as they can isolate drums, bass, vocals, and other instruments into separate stems.

1. Model Selection:
   • Primary Model: Use Demucs v4: htdemucs_ft for drum separation. This model is specifically recommended for extracting drums and bass with high quality and minimal artifacting, even in busy mixes like death metal.
   • Alternative Model: Demucs4 HT or BS Roformer SW (available via MVSEP integration in UVR) can produce six stems (vocals, drums, bass, piano, guitar, other) with excellent quality. BS Roformer SW is noted for superior multi-stem separation.
   • Ensemble Mode: Combine Demucs v4: htdemucs_ft with MVSep Drums or BS Roformer SW in ensemble mode to improve drum isolation accuracy.
2. Settings:
   • Segment Size (Chunks): Default (e.g., 32 or 64) is usually sufficient. For complex tracks with dense percussion, try a larger segment size (e.g., 128) for better quality at the cost of longer processing.
   • Overlap: Set to 0.5–0.8 for a balance of quality and speed. Higher overlap (e.g., 0.9) can improve clarity in busy mixes.
   • Output Format: Export in WAV for high-quality drum loops suitable for practice or teaching.
   • GPU Acceleration: Enable GPU rendering with an NVIDIA GPU for faster processing (e.g., 30 seconds for a 4-stem separation on an RTX 2070).
3. Post-Processing for Drum Loops:
   • DrumSep Model: For finer drum separation, use the MVSep DrumSep model, which splits drums into kick, snare, toms, and cymbals (hi-hat, ride, crash). This is ideal for creating specific drum loops or isolating elements for teaching.
   • Cleaning Artifacts: If residual bass or guitar bleed occurs, import the drum stem into a DAW and apply a high-pass filter (e.g., 100–150 Hz) to remove low-frequency bleed or use noise reduction tools.
   • Loop Creation: For practice or song assembly, use a DAW to trim the drum stem into loops. Tools like Ableton Live or Logic Pro can help quantize or adjust timing for seamless loops.

Tips for Success

• Song Complexity: Drum separation works best in tracks with distinct drum parts. In dense or heavily compressed mixes (e.g., modern pop or metal), some bleed from other instruments may occur. Test with Demucs v4 and BS Roformer SW to find the best model for each song.
• Practice Use: For teaching, export stems at 48kHz WAV to ensure clarity when demonstrating drum patterns. Use the DrumSep model to isolate specific drum components (e.g., kick or snare) for focused lessons.
• Batch Processing: UVR supports batch processing, so you can process multiple tracks at once for efficient loop creation.

3. Is UVR the Best for Stem Separation in 2025?
UVR is widely regarded as one of the best tools for vocal and drum separation in 2025 due to its open-source nature, high-quality AI models, and flexibility. However, other tools offer unique advantages, particularly in ease of use or specific use cases. Below is a comparison of UVR with top alternatives, focusing on vocal and drum separation quality, ease of use, and suitability for your needs.

Why UVR Stands Out

• Quality: UVR’s MDX-Net and Demucs v4 models deliver studio-quality vocal and drum separation, often outperforming commercial tools like iZotope RX or Spleeter in specific scenarios. Its ensemble mode enhances results by combining multiple models.
• Flexibility: UVR supports a wide range of models (e.g., MDX23C, Kim Vocal, Demucs, BS Roformer) and allows customization of settings like segment size and overlap, making it ideal for advanced users.
• Free and Open-Source: Unlike paid tools, UVR is free, with no limitations, and supports Windows, macOS, and Linux, including Apple Silicon.
• Community Support: Continuous updates and a strong community (e.g., GitHub, Discord) ensure access to the latest models and troubleshooting.
• Drum Separation: UVR’s Demucs v4 and DrumSep models are particularly effective for extracting clean drum stems, even in complex mixes, making it a top choice for your drum loop needs.

Limitations of UVR

• Learning Curve: UVR’s interface is utilitarian, and selecting the right model/settings requires experimentation, which can be intimidating for beginners.
• Processing Time: Without a compatible GPU, processing can be slow (e.g., up to an hour for complex tracks).
• No Real-Time Preview: You must process the file to hear results, which can slow workflows compared to tools with real-time previews.
• Installation: UVR requires a 1.5–5GB download and may have compatibility issues on older systems (e.g., Windows 7 or Intel Pentium CPUs).

Comparison with Alternatives
Here’s how UVR compares to other leading vocal and drum separation tools in 2025, based on available data:

1. LALAL.AI:
   • Pros:
     • Browser-based, user-friendly interface, ideal for casual users or beginners.
     • Supports vocal, drum, bass, guitar, piano, and more, with high-quality separation using Stem Splitter and Lead & Back Vocal Splitter.
     • Fast processing (e.g., 20 seconds for a 3m40s song) and supports multiple formats (MP3, WAV, FLAC, MP4).
   • Cons:
     • Subscription-based, with limited free tier (e.g., 10 minutes of processing).
     • Less customizable than UVR; fewer model options for fine-tuning.
   • Suitability: Great for quick, hassle-free vocal and drum separation, but less flexible for advanced users or those needing free tools.
2. Moises.ai:
   • Pros:
     • Browser-based and mobile-friendly, with an intuitive interface.
     • Separates vocals, drums, bass, and more, with tools to adjust tempo/pitch for practice (ideal for your teaching use case).
     • Can split lead and backing vocals, similar to UVR’s capabilities.
   • Cons:
     • Subscription-based with a limited free plan.
     • Separation quality is slightly inferior to UVR for complex tracks, with occasional artifacts.
   • Suitability: Excellent for practice and teaching due to tempo/pitch controls, but UVR outperforms for studio-quality drum and vocal stems.
3. iZotope RX 11:
   • Pros:
     • Professional-grade audio repair toolkit with Music Rebalance module for vocal and drum separation.
     • Cleaner separations with fewer artifacts than UVR in some cases, especially for noisy recordings.
     • Integrates with DAWs and offers real-time previews.
   • Cons:
     • Expensive (paid software, not open-source).
     • Less specialized for drum separation compared to UVR’s Demucs models.
   • Suitability: Best for professionals needing a complete audio repair suite, but UVR is more cost-effective for vocal/drum separation.
4. SpectraLayers 11:
   • Pros:
     • Advanced stem separation with modified models for vocals, brass, and backing vocals.
     • Improved GPU acceleration in recent updates, though still slower than UVR on NVIDIA GPUs.
   • Cons:
     • Paid software, less cost-effective than UVR.
     • Slower processing (e.g., 1 hour for a 1-hour file without CUDA).
   • Suitability: Good for users already invested in Steinberg products, but UVR offers better value and speed for vocal/drum separation.
5. MVSEP:
   • Pros:
     • Web-based service with state-of-the-art models (e.g., BS Roformer, MelBand Roformer) integrated into UVR.
     • Excellent for multi-stem separation (vocals, drums, bass, piano, guitar), with high ratings for drum and bass isolation.
   • Cons:
     • 100MB file size limit for uploads, less versatile than UVR’s standalone software.
     • Requires internet access, unlike UVR’s offline capability.
   • Suitability: Strong for drum separation, but UVR’s integration of MVSEP models makes it redundant unless you prefer a web-based workflow.
6. RipX:
   • Pros:
     • High-quality stem separation, reportedly better than iZotope RX and Acon Remix in some tests.
     • User-friendly for isolating vocals and drums.
   • Cons:
     • Paid software, less accessible than UVR.
     • Less community support and fewer model options compared to UVR.
   • Suitability: A strong alternative for paid users, but UVR’s free models often match or exceed its performance for vocals and drums.

Is UVR the Best?

• For Your Needs: UVR is likely the best choice for vocal and drum separation in 2025 due to its free, open-source nature, high-quality models (MDX23C, Demucs v4, BS Roformer), and flexibility for advanced users. Its ability to handle complex tracks and produce clean drum loops (via Demucs and DrumSep) makes it ideal for your practice, song assembly, and teaching goals. The ensemble mode and post-processing options (de-reverb, de-echo) further enhance vocal clarity.
• Alternatives for Ease of Use: If you prioritize simplicity, LALAL.AI or Moises.ai offer browser-based interfaces that are easier for beginners but lack UVR’s customization and cost-free model.
• Alternatives for Professionals: iZotope RX 11 is better for professional audio engineers needing real-time previews and a broader repair toolkit, but it’s expensive and less specialized for drum separation.
• Drum-Specific Advantage: UVR’s Demucs v4 and DrumSep models are unmatched for drum loop extraction, especially in complex genres like death metal, making it superior to most alternatives for your drum-focused use case.

Final Recommendations

• For Vocals:
  • Use MDX23C-InstVoc HQ or Kim Vocal 2 with ensemble mode (combine with Voc-FT).
  • Apply VR De-Reverb and VR De-Echo-Normal for cleaner results.
  • For lead/backing vocal separation, use MDX-B Karaoke or Medley Vox (via MVSEP integration).
  • Export in WAV and test multiple models per song for optimal results.
• For Drums:
  • Use Demucs v4: htdemucs_ft or BS Roformer SW for multi-stem separation.
  • Use MVSep DrumSep for isolating kick, snare, toms, and cymbals.
  • Set moderate segment size (32–64) and overlap (0.5–0.8), exporting in WAV.
  • Clean up bleed in a DAW if needed.
• Why Choose UVR: UVR’s combination of free access, high-quality AI models, and community support makes it the top choice for vocal and drum separation in 2025, especially for your use cases. Its Demucs v4 and DrumSep models are particularly strong for drum loops, and MDX-Net models excel for vocals. If ease of use is a priority, test LALAL.AI or Moises.ai, but UVR’s flexibility and cost-free nature are hard to beat.
• Additional Tips:
  • Download the latest UVR version (v5.6.0 or higher) from the official GitHub: https://github.com/Anjok07/ultimatevocalremovergui.
  • Use the Download Center in UVR to access “pro” models like MDX23C or Kim Vocal 2 (free with a Patreon code).
  • Join the UVR Discord community for real-time advice on model selection and troubleshooting.
  • If budget allows, consider iZotope RX 11 for professional-grade post-production, but UVR is sufficient for most of your needs.

Conclusion
UVR is an excellent choice for separating vocals and drums in 2025, offering high-quality results, flexibility, and no cost. For vocals, MDX23C-InstVoc HQ with post-processing (de-reverb, de-echo) ensures clean stems, with MDX-B Karaoke for lead/backing separation. For drums, Demucs v4: htdemucs_ft and DrumSep provide precise loop extraction, ideal for practice and teaching. While LALAL.AI and Moises.ai are easier to use, and iZotope RX 11 offers professional features, UVR’s performance and price (free) make it the best overall for your needs. Experiment with models and settings, and leverage UVR’s community for song-specific advice.

If you have further questions or need help with specific tracks, let me know, and I can guide you through tweaking settings or troubleshooting!

Sources:




20 -Beiträge



25 Websites

 

Holy Moly! PulseWave is like a human ChatGPT on steroids. I can't thank you enough. This thread should be saved!

You wrote, "If budget allows, consider iZotope RX 11 for professional-grade post-production, but UVR is sufficient for most of your needs."

Would that be for stem separation or for cleaning up the UVR stems, i.e., de-noise, de-reverb, etc.?

 

I am having trouble finding these:

MDX-B Karaoke
Medley Vox (via MVSEP integration)
MVSep DrumSep

 

iZotope RX 11 is recommended primarily for cleaning up UVR stems (e.g., de-noising, de-reverb, de-clicking, spectral repair, etc.) rather than for stem separation itself. UVR (Ultimate Vocal Remover) is already effective for stem separation, splitting audio into vocals, instruments, drums, etc. However, UVR's output can sometimes include artifacts, noise, or reverb that RX 11 can polish professionally. For most users, UVR handles separation well enough, and RX 11 would be overkill unless you need advanced post-production cleanup for high-quality results.

 

mvsep ones are exclusive of their own, they are not public

 

Is there anyway I/We can I get them?

 

no, its not public... but you can use them. on their site.... its free

 

I've been using MVSEP a lot lately, it's way faster and a lot of times better than my self hosted UVR.

 

Does anyone know which website posts the most accurate, most recent benchmark results of all the models/algorithms for various stem seps?

 

Spoiler: Source: ChatGPT KI

If you're asking: “Which website posts the most accurate, most recent benchmark results for stem separation (various models/algorithms)?” — the clear answer is:

✅ MVSEP.com (Music & Voice Separation)
• MVSEP hosts an up-to-date Quality Checker and leaderboards covering a wide range of models and model ensembles, with metrics (SDR scores) for vocals, instrumentals, bass/drums/other, etc. Reddit+4arXiv+4AudioSEX - Professional Audio Forum+4Reddit+15mvsep.com+15arXiv+15
• The leaderboard is dynamic and updated regularly as new models appear (e.g. BS Roformer, MelRoformer, SCNet XL, MDX23C, htdemucs, etc.) Reddit+10mvsep.com+10mvsep.com+10
• MVSEP metrics come from both public test sets (Musdb18HQ) and proprietary multisong/synth datasets Reddit+7mvsep.com+7music.ai+7

TL;DR Model Highlights
Best for vocals/instrumentals:
Use the Ensemble models on MVSEP which combine UVR‑MDX‑NET variants (e.g. Voc_FT, MDX23C), BS Roformer, Mel‑Band Roformer, and SCNet XL. Ensemble from late 2024 (Dec) yields highest SDR for vocals (~14.09) and instrumental (~13.79) Reddit+15mvsep.com+15mvsep.com+15.

Best for multi‑stem separations (bass, drums, other):
The ensemble using Demucs4 HT variants, Demucs3 mmi, with ensemble of vocal models gives top overall SDR average (~13.67 in mid‑2025) Reddit+14mvsep.com+14mvsep.com+14.

Best standalone single models:

• For vocals: UVR‑MDX‑NET‑Voc_FT or MDX Kimberley Jensen v2/v1 provide strong SDR on public Musdb18HQ Reddit+4mvsep.com+4mvsep.com+4.
  
  
• For bass/drums/other: Demucs3 Model B or htdemucs_ft variants are top picks arXiv+9mvsep.com+9AudioSEX - Professional Audio Forum+9.
  

️ What the community says (Reddit opinions):
“UVR5 … even better than lalalai… I keep being blown away by UVR.” Reddit+8Reddit+8Reddit+8
“UVR5 is the only answer… GOAT” Reddit
“If you still have instrumental bleeding … use MDX23C ensemble” Reddit

In summary: users consistently recommend Ultimate Vocal Remover (UVR5), especially using its ensemble mode with MDX variants and Demucs.

So: Which site to track for the most accurate & recent benchmarks?
➡️ MVSEP.com is your official, most reliable source. The Quality Checker / Leaderboard pages should be your go-to reference; they give up‑to‑date SDR metrics and highlight the top-performing models and custom ensembles arXiv+14GitHub+14Reddit+14.

If you're using UVR or MVSEP:

• Visit MVSEP’s leaderboard pages:
  
  • Leaderboard for full algorithm comparison by stem type.
    
    
  • Leaderboard2 for multi‑stem model ranking (4 vs 6 stems) Reddit+12GitHub+12mvsep.com+12mvsep.com+10AudioSEX - Professional Audio Forum+10Reddit+10.
    
• Replicate top model or ensemble in UVR’s GUI using the specified models (e.g. Voc_FT, MDX23C, BS Roformer) and settings (shifts, overlap, average averaging, etc.) GitHubScribd.
  
  
• Use Colab notebooks (e.g. MVSEP‑MDX23 Colab) if you want control or lack GPU locally (you’ll get same models used in MVSEP) Applio+1Reddit+1.
  

Summary Table
Use Case Recommended Best Source / Model
For benchmarking & comparisons MVSEP.com Quality Checker / Leaderboards
Best vocals/instrumental Latest MVSEP ensemble (MDX‑NET Voc_FT + MelRoformer + SCNet XL)
Best full multi‑stem output MVSEP ensemble including Demucs4 HT variants and vocal models
Running locally via UVR Use UVR5 GUI → replicate ensemble from leaderboard
Running via cloud or Colab MVSEP.com queue or MVSEP‑MDX23 Colab notebooks
Let me know if you'd like help parsing the leaderboard charts or replicating specific ensemble settings within UVR5 or Colab!

 

Thanks so much, PulseWave.

 

I am extremely grateful for all the advice and education regarding UVR, and I'm sure (with enough patience and experimentation, the results would be fantastic . . . however, I find this app way too complicated and convoluted. I'm trying to make music, and I'm certain going down this rabbit hole will keep me from actually making music. For many, diving deep into all these modules will be a fun adventure. In my case, using UVR is a stressful endeavor that will lead to unproductively . . . not to mention the 44.1 limitation.

With that in mind, what other affordable, straightforward options are out there that yield good results?

Just as I was about to give up on UVR, StudioJuan from Reddit wrote, "
UVR. Open Source, FREE, and THE BEST, without question. You can try the latest up-to-date models online before you download at uvronline.app The app seems intimidating at first, but it's worth the hassle. Once you get it, it's incredibly fast, and you're running everything locally on your computer. I'm happy to help if someone has questions. Not much info online on how it works, but I've become quite an expert. Might even consider making a mini video tutorial or text guide if there's any interest.

Just so you know: Most subscription online services use EXACTLY the same models. They're just an automated AI workflow behind the curtain with a nice user interface / front end on top.

And one last tip: For getting the best possible quality audio files to work with, check out "lucida (dot) to" "

 

Last edited: Aug 3, 2025

UVR stands for (Ultimate Vocal remover). Also, I would only worry about ensemble mode, if single model performance is not generating satisfactory results.

 

I'm still confused about ensemble mode, and I apologize to PulseWave for not understanding . . . Does ensemble mode take all models selected and generate a bunch of stems for each model, or are the models combined in some way to make one more "complete" stem(s). I keep reading and researching, but instead of gaining clarity I am increasing confusion. Any thoughts are appreciated. What is the simplest, most effective way to get the best stem sep results . . . in my case . . . lifting drum loops from full mixes?

Thanks in advance.

 

Spoiler: Source: Grok KI

I understand your confusion about Ultimate Vocal Remover’s (UVR) Ensemble Mode, and I’m happy to clarify it for you, especially since you’re focused on extracting drum loops from full mixes. I’ll explain Ensemble Mode clearly, address how it works for stem separation, and provide the simplest, most effective approach for isolating high-quality drum stems, leveraging UVR’s capabilities and insights from available resources.

What is Ensemble Mode in UVR?
Ensemble Mode in UVR is a feature that combines the outputs of multiple AI models to produce a single, optimized stem (or pair of stems, like drums and no-drums). Here’s how it works:

• Not Multiple Stems per Model: Ensemble Mode does not generate separate stems for each selected model. Instead, it processes the audio through each chosen model and then combines their outputs using a specified algorithm (e.g., Max Spec, Min Spec, or Avg/Avg) to create one final stem (or stem pair) that aims to be more accurate and cleaner than any single model’s output.
• How Models Are Combined: Each model in UVR (e.g., Demucs v4, MDX-Net, VR Arch) has strengths and weaknesses due to differences in training data, architecture, or focus (e.g., vocals vs. drums). Ensemble Mode leverages these differences by running the audio through all selected models and then using a mathematical approach to merge their results. For example:
  • Max Spec/Min Spec: Selects the maximum or minimum spectral components from the models’ outputs to enhance clarity and reduce artifacts.
  • Avg/Avg: Averages the outputs to balance quality and minimize errors.
  • This combination reduces bleed (unwanted elements like vocals in a drum stem) and artifacts by emphasizing the best parts of each model’s separation.
• Why Use Ensemble Mode?: It’s designed to improve separation quality by compensating for individual model limitations. For instance, one model might excel at isolating drum transients but struggle with bass bleed, while another handles bass better. Ensemble Mode blends these strengths for a more “complete” stem.

This is supported by discussions on UVR’s GitHub and Reddit, where users note that Ensemble Mode “combines various processing models to give the best possible result” by leveraging the strengths of each model.

Simplest, Most Effective Way to Extract Drum Loops with UVR
Since your goal is to lift drum loops from full mixes, I’ll outline a straightforward approach using UVR 5 (version 5.6.0 as of the latest info) to achieve high-quality drum stems with minimal complexity. This approach prioritizes ease of use, quality, and performance, tailored to your needs as a purist seeking clean drum loops for use in a synth like Harmor or another high-quality engine.

Step-by-Step Guide

1. Install UVR 5:
   • Download UVR 5 from its official GitHub page. It’s free, open-source, and available for Windows, macOS, and Linux.
   • Ensure you have FFmpeg installed in the UVR directory for audio format support, as mentioned in the setup guide.
   • For performance, UVR can use GPU (NVIDIA RTX 1060 or better recommended) or CPU. GPU is faster (e.g., 30 seconds for a 4-stem separation on an RTX 2070 Super), but CPU works fine for low-end systems.
2. Select Input and Output:
   • Open UVR 5 and choose your input audio file (supports MP3, WAV, FLAC, etc., up to 350MB). Select an output folder and format (WAV for highest quality).
3. Choose Process Method and Models:
   • Process Method: Select “Demucs” from the dropdown menu, as Demucs models (especially v4) are optimized for multi-stem separation, including drums, and consistently deliver “impeccable results” for drum stems.
   • Main Stem Pair: Choose “Drums/No Drums” to isolate the drum track and a complementary “no drums” track (containing vocals, bass, etc.).
   • Enable Ensemble Mode:
     • In Ensemble Mode, select multiple Demucs models for drums. Recommended models (available in UVR’s Download Center):
       • Demucs v4: htdemucs_ft: Fine-tuned for drums and bass, noted for excellent drum separation.
       • Demucs v4: htdemucs: General-purpose 4-stem model (vocals, drums, bass, other), reliable for drums.
       • MVSep Drums: Specialized for drum separation, high-quality results.
     • Download these models from the Settings > Download Center tab.
     • Ensemble Algorithm: Choose Max Spec for higher quality (takes longer) or Avg/Avg for a balance of quality and speed. Max Spec is recommended for cleaner drum stems.
4. Adjust Settings (Keep Simple):
   • Segment Size: Controls how much audio is processed at once. Default (e.g., 32 or 64) is fine for most systems. Larger segments improve quality but increase processing time and resource use. For low-end PCs, stick to defaults.
   • Overlap: Higher overlap (e.g., 0.75–1.0) improves quality by smoothing transitions between segments but increases processing time. Start with 0.75.
   • Shifts: Multiple passes (e.g., 2–3 shifts) can reduce artifacts but increase processing time. For simplicity, set to 1 or 2.
   • Denoise: Set to “None” or “Standard” to avoid over-processing drum transients, which can sound metallic.
5. Process and Review:
   • Click “Start Processing.” For a 3–4 minute song, expect 30 seconds to a few minutes with a GPU, or longer with a CPU.
   • UVR will output two files: a drum stem and a “no drums” stem. Check the drum stem for clarity and minimal bleed (e.g., vocals or bass leaking into the drums).
   • Use UVR’s “Sample Mode” to process a 30-second clip first to test settings before committing to the full track.
6. Optional Refinement:
   • If the drum stem has artifacts or bleed, try adding a VR Arch model (e.g., MVSep Drums) to the ensemble for better vocal/backing vocal removal.
   • For complex mixes (e.g., death metal), adjust VR Arch settings like Window Size (1024) and Aggression Level (1) to fine-tune drum extraction.
   • If bleed persists, use a multi-pass approach: Process the “no drums” stem again with a vocal-focused model (e.g., MDX-Net: UVR-MDX-NET Inst Main) to isolate and subtract residual vocals, then combine with the drum stem via phase inversion.

Recommended Ensemble for Drums
For the simplest, most effective drum separation:

• Models: Demucs v4: htdemucs_ft + Demucs v4: htdemucs + MVSep Drums.
• Ensemble Algorithm: Max Spec.
• Settings: Overlap = 0.75, Shifts = 2, Segment Size = Default, Denoise = None. This combination leverages Demucs’ strength in drum separation and MVSep’s specialized drum model, with Max Spec ensuring high-quality output. Users on Reddit and AudioSEX report excellent drum stem results with Demucs v4, often outperforming paid tools like iZotope RX or RipX.

Tips for Best Results

• Source Material Matters: Drum separation quality depends on the mix. Clear, well-separated tracks (e.g., modern pop or electronic) yield better results than dense or muddy mixes (e.g., death metal).
• Experimentation: No single model or setting is universally best. Test different ensembles and tweak overlap/shifts if the initial results have artifacts.
• Post-Processing: If minor artifacts remain, use a DAW (e.g., FL Studio) to apply EQ or gating to clean up the drum stem. For Harmor integration, import the drum stem and use its additive synthesis to enhance transients or reshape harmonics.
• Avoid Overcomplicating: Stick to 2–3 models in Ensemble Mode to keep processing manageable. More models increase time without guaranteed improvements.
• Community Resources: Check the Audio Separation Discord or UVR’s Google Doc for model recommendations and updates.

Why This Works for Your Goal

• Drum Focus: Demucs v4 and MVSep Drums are trained to isolate drums, capturing kicks, snares, and cymbals with minimal bleed, ideal for creating clean loops for Harmor or other synths.
• High-Quality Engine: Unlike Harmor’s “bad spectral quality” in resynthesis, UVR’s Demucs models operate at full bandwidth (no 11kHz cutoff like Spleeter), ensuring crisp, high-fidelity drum stems.
• Performance: The recommended settings balance quality and speed, suitable for low-end PCs, and avoid resource-heavy upscalers, aligning with your preferences.

Addressing Your Confusion
Your confusion likely stems from the misconception that Ensemble Mode outputs multiple stems per model. Instead, it’s a “team effort” where models collaborate to produce one optimized stem. Think of it like blending multiple chefs’ recipes to make one perfect dish, not serving each chef’s dish separately. The Max Spec algorithm, for instance, picks the best spectral components from each model to reduce artifacts, as explained in UVR guides.

If you’re still unsure, could you clarify what specific issues you’re encountering (e.g., artifacts, bleed, or processing time)? This would help me refine the settings further. For now, the above setup should give you clean drum loops ready for use in a high-quality synth like Alchemy, Serum, or Vital, as discussed in my previous response.

Let me know if you need help with a specific UVR setting, integrating the drum stem into Harmor, or adapting it to another synth!

 

Um... your topic name is wrong. UVR stands for "Ultimate Vocal Remover", not "Recorder". Just saying...