What is Custom stem extract / remove?

One extracted target stem or a mix with that target removed.; Practice tracks, accompaniment, remixing, sampling, and targeted cleanup.

Who is Custom stem extract / remove for?

Separation & stem split One extracted target stem or a mix with that target removed.; Practice tracks, accompaniment, remixing, sampling, and targeted cleanup.

What tasks can Custom stem extract / remove help with?

Separation & stem split; One extracted target stem or a mix with that target removed.; Practice tracks, accompaniment, remixing, sampling, and targeted cleanup.; Music audio or video; choose vocals, drums, bass, guitar, piano, or other.; Extract only the selected instrument; Remove one target while preserving the remaining mix

Custom stem extract / remove · TelkNet

Separation & stem splitGPU workload

Custom stem extract / remove

Q: Which model and keyword entities are connected with Custom stem extract / remove?

BS-RoFormer SW Fixed 6-stem, MVSep BS Roformer SW, MVSep Multisong leaderboard, BS-Rofo-SW-Fixed HF weights, BS-Rofo-SW-Fixed mirror, BS-RoFormer implementation, bs-roformer-infer registry, python-audio-separator

Practice tracks, accompaniment, remixing, sampling, and targeted cleanup.

One extracted target stem or a mix with that target removed.

6 credits/useFile inputMax 100 MBEstimated 120-900 seconds

Input

Music audio or video; choose vocals, drums, bass, guitar, piano, or other.

Supported formats

mp3wavflacoggm4aaacaifaiffmp4movmkv

Output

One extracted target stem or a mix with that target removed.

Use case

Practice tracks, accompaniment, remixing, sampling, and targeted cleanup.

Models, parameters, and sources

Six-Stem Separation splits a mix into six WAV stems: vocals, drums, bass, guitar, piano, and other.

BS-RoFormer SW Fixed 6-stemMVSep BS Roformer SWMVSep Multisong leaderboardBS-Rofo-SW-Fixed HF weightsBS-Rofo-SW-Fixed mirrorBS-RoFormer implementationbs-roformer-infer registrypython-audio-separatorBS-RoFormerHT Demucsfacebookresearch/demucsHT Demucs paper

BS-RoFormer SW Fixed 6-stem

BS-RoFormer SW Fixed uses BS-Rofo-SW-Fixed.ckpt plus BS-Rofo-SW-Fixed.yaml to split a full mix into six WAV stems.

MVSep BS Roformer SW MVSep Multisong leaderboard BS-Rofo-SW-Fixed HF weights BS-Rofo-SW-Fixed mirror BS-RoFormer implementation bs-roformer-infer registry python-audio-separator BS-RoFormer

Audio output: vocals, drums, bass, guitar, piano, and other WAV stems.
Useful for arrangement review, practice material, remix prep, and turning a mixed song into separate stem files.

Model comparison

Model

Role

Why it is used

Output impact

ModelBS-RoFormer SW Fixed 6-stem

MVSep BS Roformer SW MVSep Multisong leaderboard BS-Rofo-SW-Fixed HF weights BS-Rofo-SW-Fixed mirror BS-RoFormer implementation bs-roformer-infer registry python-audio-separator BS-RoFormer

RoleBS-RoFormer SW Fixed six-stem audio separator

Why it is usedSelected because BS-RoFormer SW Fixed 6-stem outputs vocals, drums, bass, guitar, piano, and other, with public MVSep and BS-RoFormer references.

Output impactSix WAV stems

ModelHT Demucs

facebookresearch/demucs HT Demucs paper

RoleSeparation baseline / fallback

Why it is usedHT Demucs is a strong open-source reference; this tool uses BS-RoFormer SW Fixed 6-stem because it directly targets the six displayed stems.

Output impactContext or fallback stems

Benchmark chart: MVSep BS-RoFormer SW Fixed six-stem data / MVSep piano-stem validation rows / MVSep guitar-stem validation rows

Open-source deployable scope: these tables include only rows with verified downloadable weights or configurations and a local deployment path. SOTA and Best are relative to the verified rows shown, not hosted/API-only or commercial systems and not a universal leaderboard.

MVSep BS-RoFormer SW Fixed six-stem data

MVSep Multisong entry 8374

This keeps the deployable BS-RoFormer SW Fixed 6-stem row beside the public htdemucs_6s baseline. Nearby Test, RA, Logic, and UVR rows are variants or test runs, not the tool model.

Evaluation context	Model	Vocals SDR (dB)	Detailed metrics
MVSep entry 8374	BS-RoFormer SW Fixed 6-stemTelkNetMVSep entry 8374 This tool uses BS-RoFormer SW Fixed for vocals, drums, bass, guitar, piano, and other.	11.3019	Instrumental SDR (dB)17.5091Bass SDR (dB)14.6235Drums SDR (dB)14.1129Guitar SDR (dB)9.0500Piano SDR (dB)7.8300Other SDR (dB)8.7169

MVSep piano-stem validation rows

MVSep Piano scores

MVSep publishes separate piano-stem validation rows for six-stem-capable separators. These rows explain why the current BS-RoFormer SW Fixed model is used for piano output.

Rank	Model	Piano SDR (dB)	Detailed metrics
MVSep entry 9578	BS-RoFormer SW FixedTelkNetMVSep entry 9578	7.8031SOTA	Other SDR (dB)19.5824Best
MVSep entry 7396	Ensemble (SCNet + Mel)MVSep entry 7396	6.2054	Other SDR (dB)18.4851
MVSep entry 7395	SCNet Large (2024.09)MVSep entry 7395	5.8912	Other SDR (dB)18.1698
MVSep entry 7394	MelRoformer (viperx)MVSep entry 7394	5.7118	Other SDR (dB)17.9917
MVSep entry 7392	mdx23c (2024.09)MVSep entry 7392	5.5996	Other SDR (dB)17.8796
MVSep entry 7390	mdx23c (2023.08)MVSep entry 7390	4.7969	Other SDR (dB)17.0766

MVSep guitar-stem validation rows

MVSep Guitar scores

MVSep publishes separate guitar-stem validation rows for six-stem-capable separators. These rows provide the deployable model context missing from a single six-stem leaderboard.

Rank	Model	Guitar SDR (dB)	Detailed metrics
MVSep entry 9245	BS-RoFormer SW FixedTelkNetMVSep entry 9245	9.0120SOTA	Other SDR (dB)15.9828Best
MVSep entry 7149	Ensemble (BSRoformer + MelRoformer)MVSep entry 7149	7.5112	Other SDR (dB)14.4820
MVSep entry 6741	Ensemble (mdx23 + MelRoformer)MVSep entry 6741	7.1835	Other SDR (dB)14.1542
MVSep entry 6776	BSRoformer (viperx)MVSep entry 6776	7.1266	Other SDR (dB)13.9775
MVSep entry 6720	MelRoformer (2024.06)MVSep entry 6720	7.0211	Other SDR (dB)13.9919
MVSep entry 6740	mdx23c (2024.06)MVSep entry 6740	6.3422	Other SDR (dB)13.3130

Paper and benchmark notes

Music Source Separation with Band-Split RoPE Transformer

SDX23 music separation track - 2023

BS-RoFormer uses band-split spectrogram features plus RoPE Transformers; the SDX23 system ranked first in the music source separation track and a smaller version reported 9.80 dB average SDR on MUSDB18HQ.

arXiv:2309.02612

Hybrid Transformers for Music Source Separation

ICASSP - 2023

HT Demucs replaces core Demucs layers with cross-domain Transformer attention and reports 9.20 dB SDR with extra data and fine-tuning.

arXiv:2211.08553

Official / repositories

MVSep BS Roformer SW MVSep Multisong leaderboard BS-Rofo-SW-Fixed HF weights BS-Rofo-SW-Fixed mirror BS-RoFormer implementation bs-roformer-infer registry python-audio-separator

Papers / technical notes

BS-RoFormer

Parameter guide

Model: BS-RoFormer SW Fixed 6-stem (BS-Rofo-SW-Fixed.ckpt).
Audio outputs: vocals, drums, bass, guitar, piano, and other WAV stems.
Model options: the public MVSep algorithm reference lists this entry as sep_type=63. TelkNet runs local BS-RoFormer SW weights, not an external MVSep API job. It has no add_opt controls, so the tool uses the official default quality configuration.

Use cases

Extract only the selected instrument
Remove one target while preserving the remaining mix
Prepare practice, accompaniment, remix, or sampling material

Workflow

1Upload music and choose the target stem
2Choose Extract or Remove; vocals use the specialist vocal path and other targets use the six-stem path
3Preview or download the result, or remux it into the source video

Pre-flight checklist

Use the vocal target for the specialist vocal route
Dense overlapping instruments can still cause bleed
Test a representative clip before submitting a large batch

Model evidence in this run

Models, parameters, and sources

Six-Stem Separation splits a mix into six WAV stems: vocals, drums, bass, guitar, piano, and other.

BS-RoFormer SW Fixed 6-stem

Model

BS-Rofo-SW-Fixed.ckpt + YAML

BS-RoFormer SW Fixed uses BS-Rofo-SW-Fixed.ckpt plus BS-Rofo-SW-Fixed.yaml to split a full mix into six WAV stems.

Audio output: vocals, drums, bass, guitar, piano, and other WAV stems.
Useful for arrangement review, practice material, remix prep, and turning a mixed song into separate stem files.

MVSep BS Roformer SW MVSep Multisong leaderboard BS-Rofo-SW-Fixed HF weights BS-Rofo-SW-Fixed mirror BS-RoFormer implementation bs-roformer-infer registry python-audio-separator BS-RoFormer

Model comparison

BS-RoFormer SW Fixed 6-stem

BS-RoFormer SW Fixed six-stem audio separator

Selected because BS-RoFormer SW Fixed 6-stem outputs vocals, drums, bass, guitar, piano, and other, with public MVSep and BS-RoFormer references.

Six WAV stems

MVSep BS Roformer SW MVSep Multisong leaderboard BS-Rofo-SW-Fixed HF weights BS-Rofo-SW-Fixed mirror BS-RoFormer implementation bs-roformer-infer registry python-audio-separator BS-RoFormer

HT Demucs

Separation baseline / fallback

HT Demucs is a strong open-source reference; this tool uses BS-RoFormer SW Fixed 6-stem because it directly targets the six displayed stems.

Context or fallback stems

facebookresearch/demucs HT Demucs paper

Parameter guide

Model: BS-RoFormer SW Fixed 6-stem (BS-Rofo-SW-Fixed.ckpt).
Audio outputs: vocals, drums, bass, guitar, piano, and other WAV stems.
Model options: the public MVSep algorithm reference lists this entry as sep_type=63. TelkNet runs local BS-RoFormer SW weights, not an external MVSep API job. It has no add_opt controls, so the tool uses the official default quality configuration.

Papers / technical notes

MVSep BS Roformer SW MVSep Multisong leaderboard BS-Rofo-SW-Fixed HF weights BS-Rofo-SW-Fixed mirror BS-RoFormer implementation bs-roformer-infer registry python-audio-separator BS-RoFormer

Input

TelkNet QA — batch-a

Demo

2 audio tracks

TelkNet QA — batch-a

One extracted target stem or a mix with that target removed.

0:00 / 0:30

Guitar

0.0 dB

0.00s

vocals

0.0 dB

0.00s

0:000:050:100:150:200:250:30

Under Maintenance

This tool is under maintenance. Please try again later.

Quick feedback

Write your suggestion or request here and submit it. We listen to every voice and work hard to turn good ideas into real features.