Text & Reference-Guided Sound Separation (AudioSep-Hive)

Q: What is Text & Reference-Guided Sound Separation (AudioSep-Hive)?

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

Q: Who is Text & Reference-Guided Sound Separation (AudioSep-Hive) for?

Separation & stem split Extract an open-vocabulary sound with AudioSep-Hive, FlowSep-Hive or CLAPSep, or isolate a source that matches a 10-second reference with Banquet.; WAV

Q: What tasks can Text & Reference-Guided Sound Separation (AudioSep-Hive) help with?

Separation & stem split; WAV; Remix, practice with accompaniment and edit separate parts.; Supports inputs in formats such as mp3 / wav / flac / ogg / m4a / aac / aif / aiff / mp4 / mov / mkv; AudioSep-Hive; FlowSep-Hive

Q: Which models and related terms are associated with Text & Reference-Guided Sound Separation (AudioSep-Hive)?

AudioSep-Hive, FlowSep-Hive, CLAPSep, Banquet, Hive source (pinned), AudioSep-Hive weights (pinned), FlowSep-Hive weights (pinned), Hive paper

Models, parameters, and sources

AudioSep-HiveFlowSep-HiveCLAPSepBanquetHive source (pinned)AudioSep-Hive weights (pinned)FlowSep-Hive weights (pinned)Hive paperCLAPSep source (pinned)Official CLAPSep Space artifacts (pinned)CLAPSep paperBanquet source (pinned)

Scores apply only to the cited dataset, metric and model version. Leading this table does not mean a current world record or a result from this service. Capabilities and tests can justify prioritizing models without comparable scores.

Benchmark chart: Text & Reference-Guided Sound Separation

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

Text & Reference-Guided Sound Separation

Hive paper, main test table

Values come from the linked public source. Row labels identify the test set or checkpoint; models without a matching result are omitted.

Public sources checked: 2026-09-10

Rank	Test setting	Model	SI-SDR (dB) ↑	Detailed metrics
#1	Hive test set	AudioSep-HiveTelkNetHive paper	5.02Best	SDR (dB) ↑5.67BestLPAPS ↓3.86BestFAD ↓0.80BestCLAP-Audio ↑0.65BestCLAP-T ↑0.31BestMUSHRA ↑68.40BestOQ ↑3.34BestRec ↑4.81Pre ↑3.40BestFai ↑4.39
#2	Hive test set	AudioSep (original)Hive paper	1.58	SDR (dB) ↑2.37LPAPS ↓4.22FAD ↓0.90CLAP-Audio ↑0.57CLAP-T ↑0.26MUSHRA ↑60.90OQ ↑2.94Rec ↑4.86BestPre ↑2.96Fai ↑4.47Best

Text & Reference-Guided Sound Separation

Hive paper, main test table

The sources use different training or task settings, so compare numbers only within the stated protocol.

Public sources checked: 2026-09-10

Test setting	Model	FAD ↓	Detailed metrics
Separation output · Hive test set	AudioSep-HiveTelkNetHive paper	0.80	SDR (dB) ↑5.67SI-SDR (dB) ↑5.02LPAPS ↓3.86CLAP-Audio ↑0.65CLAP-T ↑0.31MUSHRA ↑68.40OQ ↑3.34Rec ↑4.81Pre ↑3.40Fai ↑4.39
Generative output · Hive test set	FlowSep-HiveTelkNetHive paper	0.84	LPAPS ↓4.25CLAP-Audio ↑0.61CLAP-T ↑0.19MUSHRA ↑61.80OQ ↑3.05Rec ↑4.18Pre ↑3.35Fai ↑3.70

Text & Reference-Guided Sound Separation

CLAPSep paper, Table V

Values come from the linked public source. Row labels identify the test set or checkpoint; models without a matching result are omitted.

Public sources checked: 2026-09-10

± indicates standard deviation.

Rank	Test setting	Model	SDRi (dB) ↑	Detailed metrics
#3	2 sources · 1 target	AudioSepCLAPSep paper Table V	8.60 ± 8.86	SI-SDRi (dB) ↑7.38 ± 9.48
#2	2 sources · 1 target	CLAPSep-PTelkNetCorresponding tool model: CLAPSepCLAPSep paper Table V	10.03 ± 11.27	SI-SDRi (dB) ↑8.70 ± 12.64
#1	2 sources · 1 target	CLAPSep-PNTelkNetCorresponding tool model: CLAPSepCLAPSep paper Table V	12.67 ± 7.80Best	SI-SDRi (dB) ↑11.79 ± 8.23Best
#3	3 sources · 1 target	AudioSepCLAPSep paper Table V	4.90 ± 6.13	SI-SDRi (dB) ↑4.20 ± 6.42
#2	3 sources · 1 target	CLAPSep-PTelkNetCorresponding tool model: CLAPSepCLAPSep paper Table V	6.48 ± 7.82	SI-SDRi (dB) ↑5.73 ± 8.50
#1	3 sources · 1 target	CLAPSep-PNTelkNetCorresponding tool model: CLAPSepCLAPSep paper Table V	8.96 ± 6.13Best	SI-SDRi (dB) ↑8.09 ± 6.17Best
#3	3 sources · 2 targets	AudioSepCLAPSep paper Table V	5.30 ± 5.38	SI-SDRi (dB) ↑2.96 ± 5.54
#2	3 sources · 2 targets	CLAPSep-PTelkNetCorresponding tool model: CLAPSepCLAPSep paper Table V	6.35 ± 7.36	SI-SDRi (dB) ↑4.88 ± 8.10
#1	3 sources · 2 targets	CLAPSep-PNTelkNetCorresponding tool model: CLAPSepCLAPSep paper Table V	8.90 ± 5.78Best	SI-SDRi (dB) ↑7.82 ± 6.32Best
#3	4 sources · 1 target	AudioSepCLAPSep paper Table V	3.52 ± 5.21	SI-SDRi (dB) ↑2.81 ± 5.76
#2	4 sources · 1 target	CLAPSep-PTelkNetCorresponding tool model: CLAPSepCLAPSep paper Table V	5.94 ± 7.58	SI-SDRi (dB) ↑4.85 ± 8.29
#1	4 sources · 1 target	CLAPSep-PNTelkNetCorresponding tool model: CLAPSepCLAPSep paper Table V	7.73 ± 5.79Best	SI-SDRi (dB) ↑6.80 ± 6.00Best
#3	4 sources · 2 targets	AudioSepCLAPSep paper Table V	2.37 ± 4.26	SI-SDRi (dB) ↑0.65 ± 4.32
#2	4 sources · 2 targets	CLAPSep-PTelkNetCorresponding tool model: CLAPSepCLAPSep paper Table V	2.91 ± 5.54	SI-SDRi (dB) ↑1.37 ± 5.98
#1	4 sources · 2 targets	CLAPSep-PNTelkNetCorresponding tool model: CLAPSepCLAPSep paper Table V	5.20 ± 4.63Best	SI-SDRi (dB) ↑4.01 ± 4.95Best
#3	4 sources · 3 targets	AudioSepCLAPSep paper Table V	4.20 ± 4.51	SI-SDRi (dB) ↑2.03 ± 4.27
#2	4 sources · 3 targets	CLAPSep-PTelkNetCorresponding tool model: CLAPSepCLAPSep paper Table V	5.34 ± 6.70	SI-SDRi (dB) ↑3.79 ± 7.09
#1	4 sources · 3 targets	CLAPSep-PNTelkNetCorresponding tool model: CLAPSepCLAPSep paper Table V	8.00 ± 4.87Best	SI-SDRi (dB) ↑6.73 ± 5.20Best

Paper and benchmark notes

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

ICML - 2026

Paper

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

IEEE/ACM TASLP - 2024

Paper

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

ISMIR - 2024

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

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.

Text & Reference-Guided Sound Separation

Models, parameters, and sources

AudioSep-Hive

FlowSep-Hive

CLAPSep

Banquet

Benchmark chart: Text & Reference-Guided Sound Separation

Text & Reference-Guided Sound Separation

Text & Reference-Guided Sound Separation

Text & Reference-Guided Sound Separation

Paper and benchmark notes

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

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

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

Official / repositories

Papers / technical notes

Parameter guide

Workflow

Before you submit

Quick feedback