Input-Adaptive Spectral Feature Compression by Sequence Modeling for Source Separation
arXiv - 2026
Extract clear dialogue from film and TV audio, and split the background into separate music and sound-effects tracks.
Loading...
Separate film audio into dialogue, background music and effects for dubbing and post-production mixing.
Input
Supports inputs in formats such as mp3 / wav / flac / ogg / m4a / aac / aif / aiff / mp4 / mov / mkv
Supported formats
Output
WAV
Use case
Separate film audio into dialogue, background music and effects for dubbing and post-production mixing.
Choose a model for your material and purpose, and review the examples before submitting.
Cost, time, formats and parameters depend on the selected model.
Include the tool name, steps and the problem in your feedback. Screenshots can help us locate it.
A time-frequency model for three film stems; this is TIGER DnR, distinct from the two-speaker TIGER model.
Export dialogue, music and effects as three stems for dubbing, video editing and background remixing.
A band-split network separates dialogue, music and effects; compare dialogue leakage and background detail.
Separate film and video audio into dialogue, music, and sound effects. Outputs 44.1 kHz mono audio; stereo input is downmixed to mono first.
arXiv - 2026
Extract clear dialogue from film and TV audio, and split the background into separate music and sound-effects tracks.
IEEE OJSP - 2024
Extract clear dialogue from film and TV audio, and split the background into separate music and sound-effects tracks.
arXiv - 2024
Extract clear dialogue from film and TV audio, and split the background into separate music and sound-effects tracks.
arXiv - 2026
Separate film and video audio into dialogue, music, and sound effects. Outputs 44.1 kHz mono audio; stereo input is downmixed to mono first.
CVPR - 2026
Use AV-CASS visual and audio conditions to separate speech, sound effects and music from a video. Audio-only conditioning remains available for a direct comparison.
3 output categories
Only records with verified public inference implementations and matching weights are shown; algorithms requiring no weights are also eligible. Closed services, unreleased weights and unmatched experiment versions are excluded. This is not an all-model leaderboard.
Each set of weights appears once per benchmark, using its best published result for the selected metric. An overall score comes from one evaluation. Expand to compare evaluations and settings.
Public weights do not establish a fully open-source license. Expand the data for declared weight licenses and access conditions. Undeclared licenses are marked separately.
Markers identify models or components used here. Scores and settings come from the cited tests, not a new evaluation of the complete site pipeline.
The selected model has no verified matching public result and has not been ranked.
Choose a dataset to view all metrics for each model. Rows follow the official rank or selected metric. Bold marks each column’s best comparable value, including ties. Different evaluation protocols are ranked separately; missing scores are shown as -.
| Metric rank | Model | Music↑SDR (dB) ↑ | Effects↑SDR (dB) ↑ | Dialogue↑SDR (dB) ↑ |
|---|---|---|---|---|
| 1 | bandit_v2_multi.ckptPublic weights · Inference code | 9.0271 | 10.7765 | 12.2494 |
| 2 | bandit_63_zfturbo.ckptPublic weights · Inference code | 8.2894 | 9.7179 | 11.1476 |
| 3 | bandit_57_zfturbo.ckptPublic weights · Inference code | 8.0260 | 9.2388 | 10.8392 |
| 4 | bandit_30_zfturbo.ckptPublic weights · Inference code | 8.2824 | 9.4532 | 10.7387 |
| 5 | bandit_plus.ckptPublic weights · Inference code | 6.8979 | 8.9382 | 9.8945 |
| 6 | demucs4_cdx_zfturbo_1.ckptPublic weights · Inference code | 6.5774 | 8.9141 | 9.3908 |
| 7 | demucs4_cdx_zfturbo_3.ckptPublic weights · Inference code | 6.4663 | 8.8205 | 9.3759 |
| 8 | demucs4_cdx_zfturbo_2.ckptPublic weights · Inference code | 6.4385 | 8.7004 | 9.2300 |
| 9 | bandit_last.ckptPublic weights · Inference code | 6.4951 | 8.6135 | 8.2862 |
Datasets, versions and conditions remain separate, including sources with no eligible ranking rows.
Deployable entries 9 / source rows 21 · Checked: 2026-10-25
A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.
PT2 · MVSep Ensemble (SCNet + Mel) (2024.11) · MVSep SCNet DnR (2024.11) · MVSep MelBand Roformer DnR (v1, 2024.11) · Bandit v2 (multi) · test_scheme2 · test_scheme1 · test1012 · test_4stems_dnr_0908 · Bandit Plus (no resample) · Bandit Plus (with resample) · test_scheme3
Deployable entries 2 / source rows 10
A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.
MRX-C † · - · Hybrid Demucs (v3) †‡ · - · BS-Locoformer (Small) · C1 · Band-split (64) · SFC-Locoformer (Small) · C2 · SFC-Mamba (64) · SFC-Locoformer (Small) · C3 · SFC-CA (64) · BS-Locoformer (Medium) · D1 · Band-split (64) · SFC-Locoformer (Medium) · D2 · SFC-Mamba (64) · SFC-Locoformer (Medium) † · D4 · SFC-CA (64)
Deployable entries 2 / source rows 2
Scores apply to the listed dataset, metric and model version; results from different tests are not directly comparable.
The table includes only models with public weights or configurations that can run locally.
Choose a dataset to view all metrics for each model. Rows follow the official rank or selected metric. Bold marks each column’s best comparable value, including ties. Different evaluation protocols are ranked separately; missing scores are shown as -.
Deployable entries 2 / source rows 10
A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.
MRX-C † · - · Hybrid Demucs (v3) †‡ · - · BS-Locoformer (Small) · C1 · Band-split (64) · SFC-Locoformer (Small) · C2 · SFC-Mamba (64) · SFC-Locoformer (Small) · C3 · SFC-CA (64) · BS-Locoformer (Medium) · D1 · Band-split (64) · SFC-Locoformer (Medium) · D2 · SFC-Mamba (64) · SFC-Locoformer (Medium) † · D4 · SFC-CA (64)
Values come from the linked public source. Row labels identify the test set or checkpoint; unreported scores are shown as -.
| Metric rank | Model | Test setting | Params. | Speech | Music | Sfx | Average | ||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| SNR↑ | SISDR↑ | SNR↑ | SISDR↑ | SNR↑ | SISDR↑ | SNR↑ | SISDR↑ | ||||
| 1 | SFC-Locoformer (Medium) | D3 · SFC-CA (64) | 16.0 M | 16.38 | 16.32 | 9.70 | 9.28 | 10.56 | 9.88 | 12.21 | 11.83 |
| 2 | BandIt †Public weights · Inference code | Band-split (64) | 37.0 M | 15.7 | 15.6 | 8.7 | 8.0 | 9.8 | 8.2 | 11.4 | 10.9 |
Choose a dataset to view all metrics for each model. Rows follow the official rank or selected metric. Bold marks each column’s best comparable value, including ties. Different evaluation protocols are ranked separately; missing scores are shown as -.
SI-SDR reported by the authors on 973 complete DnR v2 test tracks. Each model uses its official window and hop; the mean covers music, speech and effects. These are not TelkNet measurements and are kept separate from other datasets or training setups.
| Metric rank | Model | Test setting | Mean SI-SDR (dB) ↑ | MX SI-SDR (dB) ↑ | DX SI-SDR (dB) ↑ | FX SI-SDR (dB) ↑ | Parameters (M) |
|---|---|---|---|---|---|---|---|
| 1 | DBS-TIGER (128)Public weights · Inference code | DnR v2 · 973 · 6 s / 3 s | 10.13 | 7.78 | 15.9 | 6.7 | 1 |
| 2 | DBS-BSRNN (128)Public weights · Inference code | DnR v2 · 973 · 3 s / 1.5 s | 9.07 | 6.48 | 15.01 | 5.73 | 4.7 |