Two-speaker speech separation on the WSJ0-2mix benchmark improved from roughly 6 dB of signal gain with deep clustering in 2015 to 25.1 dB SI-SNRi in 2024 (Hershey et al., arXiv 2015; SepReformer, NeurIPS 2024), a leap that turned the ‘cocktail party problem’ from a lab curiosity into shipping hearing-aid and headset features. The catch is that clean benchmarks flatter the models: Conv-TasNet falls from 15.3 dB on clean mixtures to 8.3 dB once noise and reverberation are added (MossFormer2 paper, arXiv 2312.11825, comparison tables). The demand side is enormous, with 430 million people needing rehabilitation for disabling hearing loss (WHO fact sheet, March 2026). We aggregated data from peer-reviewed papers on arXiv, IEEE ICASSP, Interspeech and NeurIPS, Microsoft’s DNS and NOTSOFAR-1 challenges, Google Research, MERL, WHO, the FDA, MarkeTrak 2025, and the University of Washington and Columbia University research groups. For the consumer side of the same technology, see our AI audio separator statistics.
TL;DR
- Best two-speaker WSJ0-2mix result: 25.1 dB SI-SNRi (SepReformer-L with dynamic mixing, NeurIPS 2024; TF-Locoformer L, MERL 2024).
- Deep clustering delivered about 6 dB of improvement on speaker mixtures in 2015 (Hershey et al., arXiv 1508.04306).
- Conv-TasNet hit 15.3 dB SI-SNRi with 5.1M parameters and a 2 ms minimum latency (Luo and Mesgarani, IEEE/ACM TASLP 2019).
- SepReformer-B reaches 23.8 dB at 39.8 GMACs/s versus TF-GridNet’s 23.5 dB at 460.8 GMACs/s (SepReformer, NeurIPS 2024).
- SepFormer needs 35.3 GB peak training memory versus 2.0 GB for SepMamba (S) (SepMamba paper, 2024).
- Conv-TasNet scores 15.3 dB clean, 12.7 dB noisy (WHAM!) and 8.3 dB noisy plus reverberant (WHAMR!) (MossFormer2 paper, 2024).
- Separation cut LibriCSS word error rate at 40% overlap from 43.3% to 21.7% (Chen et al., Microsoft, 2020).
- Top NOTSOFAR-1 multi-channel system: 10.8% tcpWER vs a 28.3% baseline (NOTSOFAR-1 summary, 2025).
- A 2.2 MB on-device VoiceFilter-Lite model improved WER on overlapping speech by 25.1% (Google Research, 2020).
- 430 million people need rehabilitation for disabling hearing loss, rising past 700 million by 2050 (WHO, March 2026).
- Hearing aid adoption among Americans with hearing difficulty reached 51.3%, up from 46.9% in 2022 (MarkeTrak 2025).
- UW Target Speech Hearing: 21 subjects rated the enrolled voice nearly twice as clear as unfiltered audio (University of Washington, CHI 2024).
1. A Decade of Benchmark Progress on WSJ0-2mix
The WSJ0-2mix benchmark has done for speech separation what ImageNet did for vision: a fixed target that made a decade of progress measurable. The best two-speaker score climbed from about 6 dB of signal improvement in 2015 to 25.1 dB SI-SNRi in 2024, and the decibel scale is logarithmic, so each extra 3 dB roughly halves the residual interference energy. The jump that mattered most was in 2018, when Conv-TasNet abandoned spectrogram masks and worked on raw waveforms. It became the first system to beat ideal time-frequency masks, which had long been treated as a practical ceiling (Conv-TasNet, Luo and Mesgarani).
Since then the curve has flattened. Gains from 2021 to 2024 came in steps of roughly 1 dB, and the top two 2024 models tie at 25.1 dB. The field has effectively saturated its own benchmark, which is why recent papers compete on efficiency and robustness instead (sections 2 and 3).
| Metric | Value | Source |
|---|---|---|
| Deep clustering, speaker mixtures (2015) | About 6 dB improvement | Hershey et al., Deep Clustering, arXiv 2015 |
| Improved deep clustering, two speakers (2016) | 10.3 dB SDR improvement (baseline 6.0 dB) | Isik et al., arXiv 1607.02173, 2016 |
| Conv-TasNet (2018) | 15.3 dB SI-SNRi, 15.6 dB SDRi | Luo and Mesgarani, IEEE/ACM TASLP 2019 |
| DPRNN (2019) | 18.8 dB SI-SDRi | Luo, Chen and Yoshioka, ICASSP 2020 |
| SepFormer with dynamic mixing (2020) | 22.3 dB SI-SNRi | Subakan et al., ICASSP 2021 |
| TF-GridNet, no data augmentation (2022) | 23.5 dB SI-SDRi | TF-GridNet paper, arXiv 2211.12433 |
| MossFormer2 (2023) | 24.1 dB SI-SDRi | MossFormer2 paper, arXiv 2312.11825 |
| SepReformer-L and TF-Locoformer (L), both with dynamic mixing (2024) | 25.1 dB SI-SNRi | SepReformer, NeurIPS 2024; Saijo et al., MERL 2024 |
The ideal-mask comparison is what made Conv-TasNet a landmark. On the same test set, the ideal binary mask scored 13.0 dB SI-SNRi, the ideal ratio mask 12.2 dB and the Wiener-filter-like mask 13.4 dB, all below Conv-TasNet’s 15.3 dB. Human listeners agreed: Conv-TasNet output received a mean opinion score of 4.03 versus 3.51 for the ideal ratio mask.
| Metric | Value | Source |
|---|---|---|
| Ideal binary mask (IBM) | 13.0 dB SI-SNRi | Conv-TasNet paper, 2019 |
| Ideal ratio mask (IRM) | 12.2 dB SI-SNRi | Conv-TasNet paper, 2019 |
| Wiener-filter-like mask (WFM) | 13.4 dB SI-SNRi | Conv-TasNet paper, 2019 |
| Listener MOS: Conv-TasNet vs IRM | 4.03 vs 3.51 | Conv-TasNet paper, 2019 |
| WSJ0-2mix size | 30 h train, 10 h validation, 5 h test | Conv-TasNet paper, 2019 |
| Mixing range | Random SNR between -5 and 5 dB | Conv-TasNet paper, 2019 |
Context note: ‘dynamic mixing’ (DM) creates new training mixtures on the fly and adds about 1 to 2 dB. SepFormer scores 22.3 dB with DM and 20.4 dB without it, so cross-paper comparisons need to check that flag first (SepFormer, arXiv 2010.13154). Most recent available data on the 2015 to 2019 baselines comes from the original papers.
2. Efficiency: Parameters, Compute, Memory and Latency
Accuracy has stalled, so the arms race has moved to cost. SepReformer-B reaches 23.8 dB SI-SNRi with 39.8 GMACs per second of audio, while TF-GridNet needs 460.8 GMACs/s for 23.5 dB, more than an elevenfold difference in compute for a slightly better score (SepReformer, NeurIPS 2024). Parameter counts tell a similar story: DPRNN matched the state of the art in 2019 with a model 20 times smaller than the previous best system.
Memory is the hidden cost that decides who can train these models. SepFormer peaks at 35.3 GB of memory, compared with 2.0 GB for SepMamba (S), which still scores 21.2 dB (SepMamba, Technical University of Denmark and WS Audiology, 2024). For anything that runs on a phone, a headset or a hearing aid, that difference decides whether a design is shippable.
| Metric | Value | Source |
|---|---|---|
| SepReformer-T | 22.4 dB at 3.7M params, 10.4 GMACs/s | SepReformer, NeurIPS 2024 |
| SepReformer-B | 23.8 dB at 14.2M params, 39.8 GMACs/s | SepReformer, NeurIPS 2024 |
| SepReformer-L with DM | 25.1 dB at 55.3M params, 155.5 GMACs/s | SepReformer, NeurIPS 2024 |
| TF-GridNet | 23.5 dB at 14.5M params, 460.8 GMACs/s | SepReformer, NeurIPS 2024 (comparison table) |
| QDPN | 200.0M parameters | TF-GridNet paper, 2022 |
| MossFormer2 | 55.7M params, real-time factor 0.053 | MossFormer2 paper, 2023 |
| DPRNN | 2.6M params, 20x smaller than prior best | Luo et al., ICASSP 2020 |
| Conv-TasNet minimum latency | 2 ms (vs a 32 ms STFT window) | Conv-TasNet paper, 2019 |
| Metric | Value | Source |
|---|---|---|
| SepFormer peak memory | 35.30 GB | SepMamba paper, 2024 |
| MossFormer2 with DM peak memory | 12.30 GB | SepMamba paper, 2024 |
| SepMamba (S) | 21.2 dB, 7.2M params, 2.00 GB peak memory | SepMamba paper, 2024 |
| SepMamba (M) | 22.7 dB, 22M params, 3.04 GB peak memory | SepMamba paper, 2024 |
| EConv-TasNet (Sept 2026) | 18.4 dB at 3.8M params | Liu, Novatek, arXiv 2609.11342 |
| EConv-TasNet vs Conv-TasNet | 22.4% smaller, 18.9% faster inference | Liu, arXiv 2609.11342 |
Outlier note: the newest paper in this roundup, EConv-TasNet from chipmaker Novatek (September 2026), does not chase the 25 dB record at all. It targets edge deployment, adding 3.1 dB to Conv-TasNet with fewer parameters (EConv-TasNet). That is a good signal of where commercial interest sits.
3. Real-World Robustness: Noise, Reverberation and More Speakers
Clean, studio-read speech mixed at controlled levels is not a restaurant. Conv-TasNet loses 7 dB when its test mixtures gain real ambient noise and reverberation: 15.3 dB on WSJ0-2mix, 12.7 dB on WHAM! and 8.3 dB on WHAMR! (MossFormer2 comparison tables). WHAM! adds noise recorded in coffee shops, restaurants and bars in the San Francisco Bay Area, and its authors note that most separation studies used a constrained setup with low sampling rates and no background noise (WHAM!, Interspeech 2019).
Newer models close much of that gap but not all of it. The best reverberant results sit at 17 to 18.5 dB, still 6 to 7 dB behind the clean benchmark. A third speaker costs another 1.9 to 2.8 dB for the same model (MossFormer2: 24.1 to 22.2 dB; SepFormer: 22.3 to 19.5 dB), and the LibriMix authors showed that models trained on WSJ0-2mix suffer important performance drops on similar datasets, which is why LibriMix was built from LibriSpeech speech plus WHAM! noise (LibriMix, Interspeech 2020).
| Metric | Value | Source |
|---|---|---|
| Conv-TasNet: WSJ0-2mix / WHAM! / WHAMR! | 15.3 / 12.7 / 8.3 dB SI-SDRi | MossFormer2 paper, 2023 |
| SepFormer: WHAM! / WHAMR! | 16.4 / 14.0 dB | MossFormer2 paper, 2023 |
| MossFormer2: WHAM! / WHAMR! | 18.1 / 17.0 dB | MossFormer2 paper, 2023 |
| SepReformer-L with DM: WHAM! / WHAMR! | 18.5 / 17.2 dB | SepReformer, NeurIPS 2024 |
| TF-Locoformer (M): WHAMR! | 18.5 dB | Saijo et al., MERL 2024 |
| TF-GridNet, 6-channel SMS-WSJ (full pipeline) | 22.8 dB SI-SDR, PESQ 4.08 | TF-GridNet paper, 2022 |
| Metric | Value | Source |
|---|---|---|
| Three speakers, Conv-TasNet (WSJ0-3mix) | 12.7 dB SI-SNRi | Conv-TasNet paper, 2019 |
| Three speakers, SepFormer (WSJ0-3mix) | 19.5 dB SI-SNRi | SepFormer paper, 2020 |
| Three speakers, MossFormer2 (WSJ0-3mix) | 22.2 dB SI-SDRi | MossFormer2 paper, 2023 |
| Three speakers, improved deep clustering (2016) | 7.1 dB SDR improvement | Isik et al., 2016 |
| Libri2Mix: Conv-TasNet / SepFormer / MossFormer2 | 14.7 / 19.2 / 21.7 dB | MossFormer2 paper, 2023 |
| Libri2Mix: SepReformer-M / TF-Locoformer (M) | 22.1 / 22.1 dB | SepReformer, 2024; Saijo et al., 2024 |
Caveat: the benchmark itself has a flaw. A 2025 study accepted at IEEE ASRU found that noise in the WSJ0-2mix reference signals limits achievable SI-SDR and that SI-SDR was negatively correlated with perceived noisiness across test sets (Jepsen et al., arXiv 2508.14623). Part of the last decibel of ‘progress’ may be models learning to reproduce noisy references.
4. Separation for Meetings and Speech Recognition
For transcription, decibels matter less than words. On Microsoft’s LibriCSS dataset, running separation before recognition cut word error rate at 40% speaker overlap from 43.3% to 21.7%, while barely changing results on non-overlapped audio (11.8% to 8.4%) (Chen et al., Continuous Speech Separation, 2020). The same authors warn that signal-level metrics are only weakly correlated with ASR accuracy or perceived quality, which is why meeting benchmarks score words, not SI-SNR. Our AI meeting assistant statistics show how fast the market depending on this accuracy is growing.
| Metric | Value | Source |
|---|---|---|
| LibriCSS size | 10 hours, 10 sessions, overlap 0% to 40% | Chen et al., Microsoft, 2020 |
| WER at 10% overlap: no separation vs separation | 18.8% vs 11.6% | Chen et al., 2020 (utterance-wise, 7-ch) |
| WER at 20% overlap | 27.2% vs 15.8% | Chen et al., 2020 |
| WER at 30% overlap | 35.6% vs 18.7% | Chen et al., 2020 |
| WER at 40% overlap | 43.3% vs 21.7% | Chen et al., 2020 |
| WER at 40% overlap, continuous input | 40.5% vs 21.9% | Chen et al., 2020 |
| Deep clustering ASR on two-speaker mixtures | WER 89.1% reduced to 30.8% | Isik et al., 2016 |
Real meetings are harder than simulated ones. Microsoft’s NOTSOFAR-1 challenge, run inside CHiME-8, recorded meetings with 4 to 8 participants across 30 rooms. The winning single-channel system still got 22.2% of words wrong (time-constrained tcpWER), against 10.8% for the best multi-channel system (NOTSOFAR-1 summary, arXiv 2501.17304). Microphone arrays remain the single biggest lever, and the organizers found that guided source separation is still the dominant separation technique in winning pipelines.
| Metric | Value | Source |
|---|---|---|
| NOTSOFAR-1 recorded meetings | 315 unique English meetings, 30 rooms, 4-8 participants | NOTSOFAR-1 summary, 2025 |
| Simulated training data | About 1,000 hours, 15,000 real acoustic transfer functions | NOTSOFAR-1 summary, 2025 |
| Single-channel: best vs baseline tcpWER | 22.2% vs 41.4% | NOTSOFAR-1 summary, 2025 |
| Multi-channel: best vs baseline tcpWER | 10.8% vs 28.3% | NOTSOFAR-1 summary, 2025 |
| Submissions: single-channel / multi-channel track | 9 / 5 | NOTSOFAR-1 summary, 2025 |
| VoiceFilter-Lite on-device model size | 2.2 MB (quantized) | Google Research, Nov 2020 |
| VoiceFilter-Lite WER improvement, overlapping speech | 25.1% (14.7% with reverberation) | Google Research, Nov 2020 |
Context note: Google’s VoiceFilter-Lite results (2020) are the most recent available public on-device figures from a major platform. Shipping systems since then rarely publish word-level benchmarks.
5. The Hearing Health Stakes
The cocktail party problem is not academic for the people who live with it. WHO estimates 430 million people, over 5% of the world’s population, need rehabilitation for disabling hearing loss, and projects more than 700 million by 2050 (WHO Deafness and hearing loss fact sheet, March 2026). Amplification alone raises every talker in a crowded room at once, which is why DNN-based separation is now the headline feature of premium hearing aids. Our hearing aid statistics cover the device market itself.
| Metric | Value | Source |
|---|---|---|
| People needing rehabilitation for disabling hearing loss | 430 million (over 5% of world population) | WHO fact sheet, March 2026 |
| Children with disabling hearing loss | 34 million | WHO fact sheet, March 2026 |
| People with some hearing loss by 2050 | Nearly 2.5 billion | WHO fact sheet, March 2026 |
| People needing rehabilitation by 2050 | More than 700 million (1 in 10) | WHO fact sheet, March 2026 |
| Disabling hearing loss among people over 60 | Over 25% | WHO fact sheet, March 2026 |
| Annual global cost of unaddressed hearing loss | Almost US$1 trillion | WHO fact sheet, March 2026 |
| Share living in low- and middle-income countries | Nearly 80% | WHO fact sheet, March 2026 |
Adoption is rising. MarkeTrak 2025, based on 16,148 US households and 42,982 individuals surveyed in October and November 2024, found hearing aid adoption among people with self-reported hearing difficulty reached 51.3%, up from 46.9% in 2022, with overall device satisfaction at 83% (MarkeTrak 2025, Seminars in Hearing). Consumer earbuds are entering the category too: in September 2024 the FDA authorized the first over-the-counter hearing aid software, Apple’s Hearing Aid Feature for AirPods Pro (FDA press release).
| Metric | Value | Source |
|---|---|---|
| US hearing aid adoption, people with hearing difficulty | 51.3% (46.9% in 2022) | MarkeTrak 2025 |
| Self-reported hearing difficulty, US | 9.1% | MarkeTrak 2025 |
| Hearing aid owner satisfaction | 83% | MarkeTrak 2025 |
| FDA clinical study for OTC hearing aid software | 118 subjects, mild to moderate loss | FDA, Sept 12, 2024 |
| Phonak Spheric Speech Clarity 2.0 (DNN) | 10.2 dB SNR improvement in real-world scenarios | AudiologyOnline expert article, 2026 |
| Phonak Spheric Speech Clarity 2.0 | Up to 37% reduction in listening effort, 30% more efficient | AudiologyOnline expert article, 2026 |
| Mayo Clinic and Sonova study, DNN noise reduction at 10 dB SNR | +23 and +32 percentage points speech recognition | Audiology Research, 2024 |
| Same study, SNR benefit in loud noise program | 4.8 to 5.0 dB | Audiology Research, 2024 |
Caveat: manufacturer SNR claims come from vendor-run testing, and the independent Mayo Clinic and Sonova study had only 10 participants. Phonak’s figures are reported via an AudiologyOnline expert Q&A, not a peer-reviewed paper.
6. The Frontier: Choosing Which Voice to Hear
Separating every voice is only half the problem; the device also has to know which one the listener wants. The University of Washington’s Target Speech Hearing prototype enrolls a speaker after the wearer looks at them for 3 to 5 seconds, and 21 test subjects rated that voice’s clarity nearly twice as high as unfiltered audio (UW News, May 2024). A December 2025 follow-up dropped the glance: its ‘proactive hearing assistant’ detects conversation partners from turn-taking rhythm in 2 to 4 seconds of audio (UW News, December 2025).
Other cues are being tested too. Google’s 2018 Looking to Listen model used lip movements, trained on clips drawn from 100,000 YouTube videos (Google Research blog). Columbia University’s brain-controlled approach compares separated voices with neural signals: the 2019 version was tested on three neurosurgical patients (Science Advances, 2019), and a May 2026 Nature Neuroscience study reported first human tests of a real-time system (Columbia Zuckerman Institute via EurekAlert). Personalized enrollment is also standard in communications: Microsoft’s fifth Deep Noise Suppression Challenge asked models to suppress neighboring talkers using a 10 to 30 second enrollment clip (ICASSP 2023 DNS Challenge). The same speaker-aware cleanup is what keeps a Discord voice chat intelligible when a roommate is talking in the background.
| Metric | Value | Source |
|---|---|---|
| UW Target Speech Hearing: enrollment time | 3 to 5 seconds of looking at the speaker | University of Washington, CHI 2024 |
| UW Target Speech Hearing: subjects and clarity rating | 21 subjects, nearly 2x clearer than unfiltered | University of Washington, CHI 2024 |
| UW Target Speech Hearing: direction tolerance | 16-degree margin of error | University of Washington, 2024 |
| UW proactive hearing assistant: audio needed | 2 to 4 seconds | University of Washington, EMNLP 2025 |
| UW proactive hearing assistant: partners and testers | 1 to 4 partners; 11 participants rated it more than 2x more favorably | University of Washington, Dec 2025 |
| UW proactive hearing assistant: training languages | English, Mandarin, Japanese | University of Washington, Dec 2025 |
| Google Looking to Listen: training source | 100,000 videos, about 2,000 hours of clean clips | Google Research blog, April 2018 |
| Columbia brain-controlled hearing (2019) | 3 neurosurgical patients | Han et al., Science Advances, 2019 |
| Metric | Value | Source |
|---|---|---|
| DNS Challenge edition at ICASSP 2023 | 5th | Dubey et al., Microsoft, 2023 |
| DNS 2023 audio format | Fullband 48 kHz | Dubey et al., 2023 |
| DNS 2023 enrollment clip | 10 to 30 seconds of the primary speaker | Dubey et al., 2023 |
| DNS 2023 winning score gain over noisy input | 0.145 (headset) and 0.141 (speakerphone) | Dubey et al., 2023 |
| TF-Locoformer (M) on DNS 2020 enhancement | 23.3 dB SI-SNR, STOI 98.8, PESQ-WB 3.72 | Saijo et al., MERL 2024 |
Outlier note: the UW team says the proactive system struggles most when conversations get dynamic, with participants talking over one another, which is exactly the cocktail party case. Interest in noise-cancelling headphones gives these features a ready hardware base.
Summary: Speech Separation and the Cocktail Party Problem by the Numbers
| Metric | Value | Source |
|---|---|---|
| Best WSJ0-2mix score | 25.1 dB SI-SNRi | SepReformer-L, NeurIPS 2024; TF-Locoformer (L), MERL 2024 |
| Deep clustering, 2015 | About 6 dB improvement | Hershey et al., 2015 |
| Conv-TasNet, 2018 | 15.3 dB SI-SNRi, 5.1M params | Luo and Mesgarani, 2019 |
| SepFormer with / without dynamic mixing | 22.3 / 20.4 dB | SepFormer, 2021; TF-GridNet, 2022 |
| MossFormer2 | 24.1 dB, 55.7M params | MossFormer2, 2023 |
| Conv-TasNet minimum latency | 2 ms | Conv-TasNet paper, 2019 |
| TF-GridNet vs SepReformer-B compute | 460.8 vs 39.8 GMACs/s | SepReformer, NeurIPS 2024 |
| SepFormer vs SepMamba (S) peak memory | 35.3 GB vs 2.0 GB | SepMamba, 2024 |
| Conv-TasNet clean / noisy / reverberant | 15.3 / 12.7 / 8.3 dB | MossFormer2, 2023 |
| Best WHAMR! (noisy, reverberant) | 18.5 dB | TF-Locoformer (M), MERL 2024 |
| WSJ0-3mix, MossFormer2 | 22.2 dB | MossFormer2, 2023 |
| LibriCSS WER at 40% overlap, no separation vs separation | 43.3% vs 21.7% | Microsoft, Chen et al., 2020 |
| NOTSOFAR-1 best multi-channel tcpWER | 10.8% (baseline 28.3%) | NOTSOFAR-1 summary, 2025 |
| NOTSOFAR-1 best single-channel tcpWER | 22.2% (baseline 41.4%) | NOTSOFAR-1 summary, 2025 |
| VoiceFilter-Lite WER gain on overlapping speech | 25.1% at 2.2 MB | Google Research, 2020 |
| People needing hearing rehabilitation | 430 million, 700 million+ by 2050 | WHO, March 2026 |
| US hearing aid adoption among people with difficulty | 51.3% | MarkeTrak 2025 |
| Phonak DNN SNR improvement | 10.2 dB | AudiologyOnline, 2026 |
| UW Target Speech Hearing enrollment | 3 to 5 seconds, 21 subjects | University of Washington, 2024 |
| EConv-TasNet edge model | 18.4 dB at 3.8M params | Novatek, Sept 2026 |
Methodology and Sources
Every benchmark figure above was taken from the paper that reported it, or from a later paper’s comparison table where noted. We excluded figures that appeared only in search snippets, secondary summaries or vendor marketing we could not open. Leaderboard aggregators were used only to check that no higher WSJ0-2mix result had been published.
- Hershey, Chen, Le Roux, Watanabe: Deep Clustering, arXiv 1508.04306 (2015)
- Isik, Le Roux, Chen, Watanabe, Hershey: Single-Channel Multi-Speaker Separation using Deep Clustering (2016)
- Luo and Mesgarani: Conv-TasNet, IEEE/ACM TASLP (2019); full-text tables read via ar5iv
- Luo, Chen, Yoshioka: Dual-path RNN, ICASSP 2020
- Subakan et al.: Attention Is All You Need in Speech Separation (SepFormer), ICASSP 2021
- TF-GridNet: Integrating Full- and Sub-Band Modeling for Speech Separation (2022)
- MossFormer2 (December 2023)
- Separate and Reconstruct: Asymmetric Encoder-Decoder for Speech Separation (SepReformer), NeurIPS 2024
- Saijo, Wichern, Germain, Pan, Le Roux (MERL): TF-Locoformer (2024)
- Avenstrup et al. (DTU, WS Audiology): SepMamba (2024)
- Liu (Novatek): EConv-TasNet (September 2026)
- Wichern et al.: WHAM!, Interspeech 2019
- Cosentino et al.: LibriMix, Interspeech 2020
- Jepsen, Christensen, Jensen: SI-SDR with noisy references, IEEE ASRU 2025
- Chen et al. (Microsoft): Continuous Speech Separation: Dataset and Analysis (LibriCSS) (2020)
- Abramovski et al. (Microsoft): Summary of the NOTSOFAR-1 Challenge (2025); Cornell et al.: CHiME-8 DASR Challenge (2024)
- Google Research: VoiceFilter-Lite (2020) and Looking to Listen (2018)
- Dubey et al. (Microsoft): ICASSP 2023 Deep Noise Suppression Challenge
- WHO: Deafness and hearing loss fact sheet (March 2026)
- Dobyan and Kihm: MarkeTrak 2025, Seminars in Hearing
- FDA: first over-the-counter hearing aid software authorization (September 2024)
- Saoji et al. (Mayo Clinic, Sonova): DNN noise reduction and cochlear implant candidacy, Audiology Research 2024
- AudiologyOnline: How has AI evolved with Phonak Infinio devices?
- University of Washington: Target Speech Hearing (2024) and proactive hearing assistant (2025)
- Columbia University: Han et al., Science Advances (2019) and Nature Neuroscience study release via EurekAlert (May 2026)
- Data watch: Papers disagree on baseline numbers. QDPN is listed at 22.1 dB in the TF-GridNet paper but 23.6 dB in the MossFormer2 paper; Conv-TasNet on Libri2Mix appears as 14.7 dB (MossFormer2) and 12.1 dB (EConv-TasNet); SepFormer’s compute is reported as 86.9 GMACs/s (SepReformer) and 257.94 GMACs/s (SepMamba), and its parameter count as 25.7M or 26.0M. Different papers use different input lengths, sample rates and counting tools, so compare figures within one table, not across tables. SI-SNRi and SI-SDRi are used interchangeably in this literature; the 2015 and 2016 deep clustering numbers use plain SDR improvement and are not directly comparable with later SI-SNRi scores. The NOTSOFAR-1 abstract cites 280 recorded meetings, while the paper body describes 315 unique meetings split across train, development and evaluation sets. No credible public market-size estimate specific to speech separation was found, so none is included. The Conv-TasNet, deep clustering, Looking to Listen and VoiceFilter-Lite figures are older than three years and are shown as most recent available data. The Columbia 2026 release gives no quantitative results; we cite it only for its existence and venue.
Last updated: October 3, 2026. We update this roundup quarterly, and the next refresh is expected when ICASSP 2027 accepted papers and the next CHiME challenge results are published.