FullSubNet
The full-band + sub-band fusion idea that reshaped real-time denoising — beat the top DNS Challenge 2020 systems with a simple, causal design.
About
FullSubNet's insight is that two kinds of models are both right. Full-band models see global spectral context and cross-band structure but can't exploit the fact that speech stationarity is frequency-local; sub-band models process each frequency (with context bands) independently, which is great for separating stationary noise from non-stationary speech, but they miss the big picture. FullSubNet simply chains them: a pure full-band network feeds a pure sub-band network, each doing what it's good at. On the Interspeech 2020 DNS Challenge test set, the authors report it outperforming the top-ranked systems of that challenge while staying causal and real-time. The paper appeared at ICASSP 2021, and it spawned a follow-up family — FullSubNet+ (channel attention, complex spectrograms, ICASSP 2022) and Inter-SubNet (sub-band interaction, ICASSP 2023). Conceptually it's one of the most imitated designs of the decade: TF-GridNet and MossFormer-era models all echo the full/sub-band split. The honest gaps: the authors said code would be open-sourced, but I could not confirm a stable public repo at research time, and all the numbers are author-reported.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Paper (ICASSP 2021, author-reported): beats top Interspeech 2020 DNS Challenge systems on the challenge test set
- No independent benchmark replication found
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 6/10
Scored 6/10 (TIER metrics, metrics-v1). Evidence: ICASSP 2021 peer-reviewed paper reporting wins over the top DNS Challenge 2020 systems; the design is one of the most imitated in modern enhancement (FullSubNet+, Inter-SubNet, and echoes in TF-GridNet). Strengths: elegant and influential full/sub-band fusion, causal and real-time. Docked for: (1) no confirmed stable public code release despite the paper's promise — reproducibility rests on community reimplementations; (2) every number is author-reported. License: none confirmed — no verified public release.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 5.0/10
- Openness · license + docs
- 6.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · ICASSP 2021 paper (arXiv 2010.15508); full-band + sub-band fusion, causal, real-time
- · Author-reported wins over top Interspeech 2020 DNS Challenge systems
- · Follow-ups FullSubNet+ (ICASSP 2022) and Inter-SubNet (ICASSP 2023); design widely imitated (TF-GridNet era)