Einzelspuren (Basslinien) aus einer Aufnahme herausziehen (STEM)

  • Ersteller Ersteller milo
  • Erstellt am Erstellt am

Das Problem, das ich bei meinem Workflow erkenne ist die mangelnde Möglichkeit einer Einstellung beim Midi Export aus RipX.
Die Genauigkeit des Midi Exports ist aber wohl zu genau und es dichtet dann Pausen und 1. 2. Stimme von Spielartefakten sowie viel zu genaue Tonlängen, wenn das eigentlich nur leichte Unperfektheiten des Instrumentalisten sind.

Gibt es Empfehlungen den separierten Track mit einem anderen Tool für zuverlässige, „nicht zu perfekte“ MIDI-Transkriptionen?

hier das Resultat eines zu genauen Midi Exports

1760797653364.png
 

Anhänge


Songsterr - hab's nicht gegengecheckt:




1760798454482.png


Wichtig: Songsterr nutzt Youtube Links fürs Transcribieren. Die meisten Youtube Videos sind wirklich k@cke!

Ich erstelle immer selbst ein Video, und zwar so:

- Flac aus dem Originaltrack erstellen, normalisiert auf -1db
- Video mit FFMPEG erstellen. mit dieser Batch Datei (kannst irgendwas im 16:9 Format als image.jpg nutzen ... oder das weisse Bildchen was ich hier geposted habe):

set infile=NAME_DER_FLAC_DATEI_OHNE_EXTENSION
ffmpeg -loop 1 -i image.jpg -i "%infile%.flac" -c:v libx264 -acodec flac -vf "pad=ceil(iw/2)*2:ceil(ih/2)*2" -shortest "%infile%.mp4"


- Hochladen auf YouTube, visibility=unlisted
- Diesen Link fürs AI Transcribieren nutzen


image.jpg
 
Zuletzt bearbeitet:
in welchem Format spuckt Songsterr die Transkription aus?
nur Bild oder druckbar pdf oder auch midi bzw. xml?
Ich will ja mit dem Workflow in Musescore enden.
 
Gestern gefunden, ich weiss nicht, ob jemand das schon empfohlen hat:


Eigenschaften
- gratis
- gute UI
- verschiedene Algorithmen nutzbar (auch neuere: downloadbar)
- sehr hohe Qualität mit den richtigen Settings
- Für Faule: Perplexity gibt Empfehlungen für Settings. Man bekommt Perplexity Pro als PayPal Kunde für ein Jahr gratis


Ich bin begeistert!



FRAGE: What process methods does ultimate vocal remover 5 support? Explain each!

ANTWORT PERPLEXITY: Ultimate Vocal Remover 5 (UVR5) supports multiple process methods, each representing a distinct algorithmic family for separating stems (vocals, instruments, reverb, etc.). These methods determine which AI model architecture the program uses for analysis and output.

VR Architecture​

This is the traditional architecture originally used in earlier UVR versions. It relies on convolutional neural networks specifically designed for vocal/instrument separation. It includes:
  • VR Arch Single Model V5: A classic model for general-purpose separation.
  • Variants like 7_HP2 UVR target higher frequencies and broader spectral coverage, yielding smooth results in pop and rock songs.

MDX-NET​

MDX-NET is UVR5’s primary neural network method, featuring models trained on large datasets to understand complex mixes. It delivers high-quality isolation and is ideal for most songs. Key models include:
  • UVR-MDX-NET Inst 3 / Main / HQ: For instruments or full-stem separation.
  • MDX23C-InstVoc D1581: A hybrid model that separates both vocals and instruments with strong spectral precision.
  • Supports GPU conversion, providing fast processing with minimal quality trade-offs.vocalremover.easeus+1

Demucs​

Based on Facebook AI’s Demucs architecture, this method uses waveform-level (time-domain) deep learning rather than spectrogram processing. It provides naturally sounding separations, especially effective for:
  • Full stem decomposition (vocals, bass, drums, and other instruments).
  • Models like htdemucs_ft (V4) expand performance for multi-stem extraction.deepwiki+1

ONNX Dereverb​

This process method, built from MDXNetDereverb, focuses on reverb and echo reduction rather than pure separation. It’s often applied to clean up vocals after extraction—removing room reflections, echoes, and artificial reverbs from mixed recordings.deepwiki

DeEcho​

The AudioPreDeEcho models specialize in echo suppression and clarity enhancement while performing separation. They’re ideal for live or noisy recordings, cleaning artifacts introduced by reverberation or overcompression.deepwiki

Ensemble Mode​

UVR5’s Ensemble Mode blends results from multiple models to create a single composite output. Users can select multiple different process methods (for example, MDX and Demucs) and average or combine their results using ensemble algorithms like Max Spec/Max Spec to optimize isolation fidelity.reddit+1

Summary Table​

Process MethodTypePrimary UseCharacteristics
VR ArchitectureSpectrogram CNNBasic vocals/instrumentalsFast, simple, moderate quality vocalremover.easeus+1
MDX-NETNeural net (spectro)General-purpose, high qualityMost recommended for modern mixes vocalremover.easeus+1
DemucsTime-domain DNNMulti-stem splittingMore natural sound, slower vocalremover.easeus+1
ONNX DereverbNeural dereverberatorReverb/echo cleanupPost-processing enhancer deepwiki
DeEchoNeural echo suppressorEcho removal + separationOptimized for noisy sources deepwiki
Ensemble ModeHybridBlended resultsCombines multiple models for balance vocalremover.easeus+1
In short, UVR5 lets you choose from spectral, time-domain, dereverb, and ensemble processing paths, enabling users to tailor the method to their source quality and artistic intent.
  1. https://vocalremover.easeus.com/ai-article/how-to-use-ultimate-vocal-remover.html
  2. https://www.toolify.ai/ai-news/ai-music-production-ultimate-vocal-remover-5-deep-dive-3705266
  3. https://deepwiki.com/RVC-Project/Re...sion-WebUI/3.3-uvr5-(audio-source-separation)
  4. https://www.reddit.com/r/IsolatedTr.../ultimate_vocal_remover_5_settings_questions/
  5. https://github.com/Anjok07/ultimatevocalremovergui/discussions/1222
  6. https://github.com/Anjok07/ultimatevocalremovergui/discussions/689
  7. https://www.pluginboutique.com/arti...emover-Plugin-Guide-Separate-Music-Like-a-Pro
  8. https://www.reddit.com/r/audioengin...ch_is_actually_the_best_vocal_remover_of_all/
  9. https://s3sound.com/blog/how-to-split-stems-with-uvr5
  10. https://www.reddit.com/r/IsolatedTracks/comments/vuavwq/ultimate_vocal_remover/
  11. https://ultimatevocalremover.com
  12. https://github.com/Anjok07/ultimatevocalremovergui
  13. https://www.lalal.ai
  14. https://audiosex.pro/threads/uvr-ultimate-vocal-remover-the-question-the-journey.80898/
  15. https://www.runcomfy.com/comfyui-nodes/ComfyUI-UVR5
  16. https://recording.de/threads/ultimate-vocal-remover-gui-v5-6-free.258520/
  17. https://docs.applio.org/applio/guides/audio-isolating/uvr
  18. https://vocalremover.org
  19. https://www.reddit.com/r/audioengin...ltimate_vocal_remover_5_best_settings_to_use/




FRAGE: What demucs models does does ultimate vocal remover 5 support? Explain each!

ANTWORT PERPLEXITY:
Ultimate Vocal Remover 5 (UVR5) includes support for several versions of Demucs, specializing in high-quality multi-stem separation. Each Demucs model version brings different capabilities and optimizations suited for various separation tasks. Here are the Demucs models supported and what makes each unique:

Demucs v1 and v2​

Early versions of Demucs use a time-domain, wave-to-wave neural network approach. They are relatively lightweight and can separate stems into vocals, drums, bass, and other instruments, but generally offer lower separation fidelity compared to later versions.vocalremover.easeus

Demucs v3​

Demucs v3 introduces improved neural architecture and larger training datasets. Separation quality increases, particularly for drum and bass stems, with fewer artifacts and better handling of complex mixes. It is more CPU- and GPU-intensive but produces cleaner, more natural sounding results.vocalremover.easeus

Demucs v4 (htdemucs)​

The most widely used variant in UVR5 is Demucs v4, also called htdemucs (hybrid time-frequency Demucs). It combines time-domain processing with spectrogram information, dramatically improving separation for all stems:

  • htdemucs_ft: A fine-tuned version, even more precise at instrument and vocal isolation, especially for drum/bass extraction.audiosex
  • htdemucs_6s: Version for 6-stem separation (Vocals, Drums, Bass, Piano, Other, Guitar), offering more granular instrument isolation, though slightly more prone to "bleed" between stems than the 4-stem version.github+1
  • hdemucs_mmi: Multi-model inference for added consistency, usually recommended for nuanced splits.

Demucs 4-Stem and 6-Stem Modes​

  • Demucs 4-stem (Vocals, Drums, Bass, Other): The default, providing balanced performance with minimal musical residue.s3sound
  • Demucs 6-stem (Vocals, Drums, Bass, Piano, Other, Guitar): Used for projects needing more detailed breakdown, though results can vary by song and genre.gearspace

Usage in UVR5​

  • Users can select from these Demucs models in UVR5 under the process method dropdown, and different models may be preferable depending on whether the focus is on vocal clarity, instrument isolation, or overall naturalness.
  • Demucs models excel when used alone for quick separation, or within ensembles for optimal fidelity when combined with other model types.s3sound+2
In summary, UVR5 supports Demucs v1, v2, v3, v4 (htdemucs/htdemucs_ft), hdemucs_mmi, and both 4-stem and 6-stem Demucs variants, each providing different balances of speed, CPU/GPU demand, and separation accuracy.

  1. https://vocalremover.easeus.com/ai-article/how-to-use-ultimate-vocal-remover.html
  2. https://audiosex.pro/threads/uvr-ultimate-vocal-remover-the-question-the-journey.80898/
  3. https://github.com/Anjok07/ultimatevocalremovergui/issues/1942
  4. https://gearspace.com/board/electro...4-help-needed-uvr-ultimate-voice-remover.html
  5. https://s3sound.com/blog/how-to-split-stems-with-uvr5
  6. https://www.reddit.com/r/IsolatedTr.../ultimate_vocal_remover_5_settings_questions/
  7. https://github.com/Anjok07/ultimatevocalremovergui
  8. https://github.com/Eddycrack864/Ultimate-Vocal-Remover-5.6-for-Google-Colab
  9. https://www.reddit.com/r/audioengin...ch_is_actually_the_best_vocal_remover_of_all/
  10. https://huggingface.co/spaces/Eddycrack864/UVR5/blob/main/gui_data/constants.py
  11. https://comfy.icu/extension/set-soft__AudioSeparation
  12. https://github.com/Anjok07/ultimatevocalremovergui/issues/344
  13. https://pypi.org/project/demucs/3.0.2/
  14. https://huggingface.co/spaces/heath...e386f5154a1456402f97bf7/gui_data/constants.py
  15. https://www.runcomfy.com/comfyui-nodes/ComfyUI-UVR5
  16. https://github.com/Anjok07/ultimatevocalremovergui/issues/1718
  17. https://ultimatevocalremover.com
  18. https://www.scribd.com/document/661...on-Mastering-Guide-UVR-MDX-Demucs-GSEP-Others
  19. https://dev.to/lalalai/best-ai-vocal-removers-in-2025-1036
  20. https://github.com/Anjok07/ultimatevocalremovergui/discussions/444
 
Zuletzt bearbeitet:
Hallo @Whying_Dutchman, vielen Dank - das schaue ich mir noch genauer an.

Ich habe inzwischen mit Hilfe von meinem CoPiloten und seinem Cousin ChatGPT nach Alternativen gesucht.

Option 1 (Empfehlung CoPilot) - Reaper, gepimpt mit Reapack, Umwandlung Audio zu Midi mit Reatune
Option 2 (Empfehlung ChatGPT) - Widi WIDI 4.60

Wir, also mein Freund CoPilot und sein Cousin ChatGPT, haben als Hauptproblem meines Workflows Midi Export RipX zu Musescore die fehlenden Einstellmöglichkeiten der Midi Generierung identifiziert. Beide empfohlenen Wege funktionieren, aber mit den bisherigen Einstellungen meinerseits bisher noch schlechter, als der Midi Export aus RipX.

Leider fehlt mir jetzt die Zeit weiter an den Einstelloptionen zu feilen, da mich momentan die Arbeit zeitlich auffrisst.

Ich werde weiter berichten.
 
Zuletzt bearbeitet:

Zurück
Oben Unten