Prompted sound separation · AudioSep-Hive / FlowSep-Hive · TelkNet

Separation & stem splitGPU workloadCLAPSep · Source bert-base-uncased · Weights roberta-base · Weights bart-base · Weights CLAPSep · Weights CLAPSep · Paper

Prompted sound separation

Describe a sound or provide reference audio to extract it from a recording and download it.

25 credits/useFile inputMax 100 MBEstimated 30-3600 seconds

Input

Supports inputs in formats such as mp3 / wav / flac / ogg / m4a / aac / aif / aiff / mp4 / mov / mkv

Supported formats

mp3wavflacoggm4aaacaifaiffmp4movmkv

Output

WAV

Use case

Remix, practice with accompaniment and edit separate parts.

Choose a model family

Prompted sound separation · AudioSep-HiveAvailableDescribe the target sound in text and extract it from a mixture; useful when no reference audio is available.Prompted sound separation · FlowSep-HiveAvailableSpecify a target in text using flow-matching separation; compare missed sounds and leakage with AudioSep on the same excerpt.Prompted sound separation · CLAPSepAvailableUse positive/negative text or reference audio to specify a target, including similar sounds to exclude.Prompted sound separation · BanquetAvailableSpecify a timbre with a 10-second reference; useful for sounds that are hard to describe but have a clean example.Prompted sound separation · SAM AudioUnder MaintenanceExtract sounds with text prompts; Base and Large offer different model sizes for comparison on the same target.

Choose a model for your material and purpose, and review the examples before submitting.

Cost, time, formats and parameters depend on the selected model.

Quick feedback

Include the tool name, steps and the problem in your feedback. Screenshots can help us locate it.

Models, parameters, and sources

CLAPSep: Remix, practice with accompaniment and edit separate parts.

AudioSep-HiveHive paper, main test tableAudioSep (original)FlowSep-HiveCLAPSep paper, Table VAudioSepCLAPSep-PCLAPSep-PNA Semantically Consistent Dataset for Data-Efficient Query-Based Universal Sound SeparationCLAPSep: Leveraging Contrastive Pre-trained Model for Multi-Modal Query-Conditioned Target Sound ExtractionA Stem-Agnostic Single-Decoder System for Music Source Separation Beyond Four StemsCLAPSep

AudioSep-Hive

Describe the target sound in text and extract it from a mixture; useful when no reference audio is available.

Hive source (pinned)AudioSep-Hive weights (pinned)

FlowSep-Hive

Specify a target in text using flow-matching separation; compare missed sounds and leakage with AudioSep on the same excerpt.

Hive source (pinned)FlowSep-Hive weights (pinned)Hive paper

CLAPSep

CLAPSep source (pinned)Official CLAPSep Space artifacts (pinned)CLAPSep paper

Banquet

Banquet: Remix, practice with accompaniment and edit separate parts.

Banquet source (pinned)Official Banquet checkpoint Banquet paper

SAM Audio Base

SAM Audio Base: Remix, practice with accompaniment and edit separate parts.

SAM Audio source SAM Audio Base weights SAM Audio Large weights

SAM Audio Large

SAM Audio Large: Remix, practice with accompaniment and edit separate parts.

SAM Audio source SAM Audio Base weights SAM Audio Large weights

Paper and benchmark notes

A Semantically Consistent Dataset for Data-Efficient Query-Based Universal Sound Separation

ICML - 2026

Extract an open-vocabulary sound with AudioSep-Hive, FlowSep-Hive or CLAPSep, or isolate a source that matches a 10-second reference with Banquet.

Paper

CLAPSep: Leveraging Contrastive Pre-trained Model for Multi-Modal Query-Conditioned Target Sound Extraction

IEEE/ACM TASLP - 2024

Extract an open-vocabulary sound with AudioSep-Hive, FlowSep-Hive or CLAPSep, or isolate a source that matches a 10-second reference with Banquet.

Paper

A Stem-Agnostic Single-Decoder System for Music Source Separation Beyond Four Stems

ISMIR - 2024

Extract an open-vocabulary sound with AudioSep-Hive, FlowSep-Hive or CLAPSep, or isolate a source that matches a 10-second reference with Banquet.

Paper

Official / repositories

Hive source (pinned)AudioSep-Hive weights (pinned)FlowSep-Hive weights (pinned)CLAPSep source (pinned)Official CLAPSep Space artifacts (pinned)Banquet source (pinned)Official Banquet checkpoint SAM Audio source SAM Audio Base weights SAM Audio Large weights

Papers / technical notes

Hive paper CLAPSep paper Banquet paper

Parameter guide

Describe the sound to extract in plain language. AudioSep-Hive and FlowSep-Hive require this field.
Choose positive text, reference audio, or both for CLAPSep.
Describe the target sound that CLAPSep should keep.
Optionally describe a sound that CLAPSep should suppress.
Upload reference audio that exemplifies the target sound.
Optionally upload reference audio for a sound to suppress.
Upload Banquet reference audio. The official contract uses a 10-second query excerpt.
FlowSep-Hive generation steps. The official default is 20; more steps take longer.
FlowSep-Hive classifier-free guidance scale. The official default is 1.0.
Banquet query-band batch size. The official inference default is 12.

Workflow

1Upload a recording and describe the target sound; for Banquet, also provide a 10-second reference.
2Run the selected separator and download the extracted sound.

What to check

Describe the sound to extract in plain language. AudioSep-Hive and FlowSep-Hive require this field.
Choose positive text, reference audio, or both for CLAPSep.
Upload Banquet reference audio. The official contract uses a 10-second query excerpt.

Models in this field

9 entries

Includes public test results for models used here and elsewhere. The complete source table is shown by default; conditions and metrics can also be selected for comparison. Scores from different test conditions are not directly comparable.

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.

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.

BenchmarkRanking category

SAM Audio · Text-prompted benchmark

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 -.

OVR is the paper’s pairwise subjective rating and depends on the comparator. Do not pool it with other experiments; it is not a retest of this site’s weights.

Rank by metric

SAM Audio · Text-prompted benchmark
Metric rank	Model	Effects↑OVR (1–5)	Dialogue↑OVR (1–5)	Speaker↑OVR (1–5)	Music↑OVR (1–5)	Instruments in the wild↑OVR (1–5)	Studio instruments↑OVR (1–5)
1	SAM Audio Public weights · Inference code Public weights Inference codeWeight license not declared	3.59	4.29	4.15	4.05	4	4.45
2	SoloAudio Public weights · Inference code Public weights Inference codeWeight license not declared	2.97	3.32	2.45	2.47	2.71	2.3
3	CLAPSepUsed on this site Public weights · Inference code Public weights Inference codeWeight license not declared	2.92	2.47	2.79	2.97	2.81	2.56
4	AudioSep Public weights · Inference code Public weights Inference codeWeight license not declared	2.88	2.85	2.79	3.51	2.59	2.45
5	FlowSep Public weights · Inference code Public weights Inference codeWeight license not declared	2.65	2.14	2.13	2.9	2.69	2.02
-	MossFormer2 Public weights · Inference code Public weights Inference codeWeight license not declared	-	-	2.54	-	-	-
-	Tiger Public weights · Inference code Public weights Inference codeWeight license not declared	-	-	2.5	-	-	-
-	Demucs Public weights · Inference code Public weights Inference codeWeight license not declared	-	-	-	-	-	4.26
-	Spleeter Public weights · Inference code Public weights Inference codeWeight license not declared	-	-	-	-	-	3.9

Original records (9)

Markers identify models or components used here. Scores and settings come from the cited tests, not a new evaluation of the complete site pipeline.

SAM Audio · Table 10 ↗Checked: 2026-09-27

All evaluation sources (2)

Datasets, versions and conditions remain separate, including sources with no eligible ranking rows.

Hive paper, main test table ↗
Deployable entries 3 / source rows 11
Not ranked: 8
A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.
LASS-Net ( Liu et al., 2022 ) · Discriminative Models · CLIPSep ( Dong et al., 2023 ) · Discriminative Models · OmniSep ( Cheng et al., 2025 ) · Discriminative Models · ZETA ( Manor and Michaeli, 2024 ) · Generative Models · FlowSep ( Yuan et al., 2025 ) · Generative Models · ZeroSep ( Huang et al., 2025 ) · Generative Models · DGMO ( Lee et al., 2025b ) · Generative Models · SAM-Audio ( Shi et al., 2025 ) · Generative Models
CLAPSep paper, Table V ↗
Deployable entries 3 / source rows 3

All original results and sources

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.

Hive paper, main test table

Deployable entries 3 / source rows 11

Not ranked: 8

A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.

LASS-Net ( Liu et al., 2022 ) · Discriminative Models · CLIPSep ( Dong et al., 2023 ) · Discriminative Models · OmniSep ( Cheng et al., 2025 ) · Discriminative Models · ZETA ( Manor and Michaeli, 2024 ) · Generative Models · FlowSep ( Yuan et al., 2025 ) · Generative Models · ZeroSep ( Huang et al., 2025 ) · Generative Models · DGMO ( Lee et al., 2025b ) · Generative Models · SAM-Audio ( Shi et al., 2025 ) · Generative Models

Values come from the linked public source. Row labels identify the test set or checkpoint; unreported scores are shown as -.

Rank by metric

Hive paper, main test table
Metric rank	Model	Test setting	Signal Fidelity ↑		Perceptual & Semantic Quality					Reference-free Quality ↑
Metric rank	Model	Test setting	SDR	SI-SDR	LPAPS ↓	FAD ↓	CLAP-A ↑	CLAP-T ↑	MUSHRA ↑	OQ	Rec	Pre	Fai
1	AudioSep (Hive) Public weights · Inference code Inference codeWeight license not declared	Discriminative Models	5.67	5.02	3.86	0.80	0.65	0.31	68.4	3.34	4.81	3.40	4.39
2	AudioSep ( Liu et al., 2024 ) Public weights · Inference code Inference codeWeight license not declared	Discriminative Models	2.37	1.58	4.22	0.90	0.57	0.26	60.9	2.94	4.86	2.96	4.47
-	FlowSep (Hive) Public weights · Inference code Inference code Weight license (repository declaration): apache-2.0	Generative Models	-	-	4.25	0.84	0.61	0.19	61.8	3.05	4.18	3.35	3.70

Original records (3)

Hive paper, main test table ↗Checked: 2026-10-25

The table includes only models with public weights or configurations that can run locally.

CLAPSep paper, Table V

Deployable entries 3 / source rows 3

Values come from the linked public source. Row labels identify the test set or checkpoint; unreported scores are shown as -.

Rank by metric

CLAPSep paper, Table V
Metric rank	Model	2-sources separation · 1-target		3-sources separation · 1-target		3-sources separation · 2-target		4-sources separation · 1-target		4-sources separation · 2-taget		4-sources separation · 3-target
Metric rank	Model	SDRi↑	SISDRi↑	SDRi↑	SISDRi↑	SDRi↑	SISDRi↑	SDRi↑	SISDRi↑	SDRi↑	SISDRi↑	SDRi↑	SISDRi↑
1	CLAPSep-PN Public weights · Inference code Inference codeWeight license not declared	12.67±7.80	11.79±8.23	8.96±6.13	8.09±6.17	8.90±5.78	7.82±6.32	7.73±5.79	6.80±6.00	5.20±4.63	4.01±4.95	8.00±4.87	6.73±5.20
2	CLAPSep-P Public weights · Inference code Inference codeWeight license not declared	10.03±11.27	8.70±12.64	6.48±7.82	5.73±8.50	6.35±7.36	4.88±8.10	5.94±7.58	4.85±8.29	2.91±5.54	1.37±5.98	5.34±6.70	3.79±7.09
3	AudioSep Public weights · Inference code Inference codeWeight license not declared	8.60±8.86	7.38±9.48	4.90±6.13	4.20±6.42	5.30±5.38	2.96±5.54	3.52±5.21	2.81±5.76	2.37±4.26	0.65±4.32	4.20±4.51	2.03±4.27

Original records (3)

CLAPSep paper, Table V ↗Checked: 2026-10-25