List your company

ResearchFormerly WizWand

Speech Separation

Benchmarks

DatasetSOTA methodMetricTrendResultsLast updated

VoxCeleb2+DCASE 1 speaker+noise (test)

11.66SISDR

4

Feb 26, 2026

LRS2-2Mix Variation Table 8 (test)

85.3Accuracy

4

Feb 26, 2026

WSJ0-5mix (test)

Conditional TasNet*

11.7SI-SNRi

4

Feb 26, 2026

WSJ0 closed speaker set (test)

6.54SDR Improvement (dB)

4

Feb 26, 2026

LibriCSS simulated (test)

TF-CorrNet (Conformer)

11.75SDRi

3

Apr 1, 2026

Real overlapping-speech recordings (test)

3.86MOS

3

Mar 12, 2026

WSJ0-2mix 25 selected utterances (test)

4.23MOS

3

Feb 26, 2026

WSJ0 3-speaker mixture (closed speaker set)

4.42SDR Improvement (dB)

3

Feb 26, 2026

MMSpeech 3mix-noisy (test)

11.51SI-SDRi

2

Apr 28, 2026

MMSpeech 2mix-noisy (test)

14.96SI-SDRi

2

Apr 28, 2026

Libri3Mix simulated (test)

15SI-SNRi

2

Feb 26, 2026

8.39SDRi (dB)

2

Feb 26, 2026

WSJ0 3-speaker mixture (open speaker set)

Deep Clustering

2.81SDR Improvement (dB)

2

Feb 26, 2026

Libri2Mix min 8 kHz

20.4SDR

1

Jun 30, 2026

SuDoRM-RF++ (SEP16-v2)

16Latency (ms)

1

Jun 4, 2026

4.26SI-SDRi (dB)

1

Feb 26, 2026

5.66SI-SDRi (dB)

1

Feb 26, 2026

GRID corpus (test)

—

—Primary metric

0

Feb 26, 2026