Uformer
U-Net meets dual-path conformer: joint speech enhancement and dereverberation in complex and magnitude domains, topping the DNS Challenge 2021 blind set.
About
Uformer fuses two successful ideas. The skeleton is a U-Net encoder-decoder, but the bottleneck processing is a dilated complex-and-real dual-path conformer: time attention and dilated convolutions capture local and global context, frequency attention models the spectral dimension, and a hybrid encoder-decoder models the complex spectrum and the magnitude simultaneously so the two domains inform each other, with encoder-decoder attention on top. On the Interspeech 2021 DNS Challenge blind test set the authors report 3.6032 DNSMOS — ahead of every other complex-domain entry including the challenge's best model — and the best MOS among SOTA complex-domain models. The paper was accepted at ICASSP 2022 with code promised at a public repo. The honest wrinkle: I could not confirm that repo at research time, so unlike DeepFilterNet or SGMSE this one is currently a paper-only model in practice, with community ports carrying the torch. Like CMGAN it is non-causal, so offline use only.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Paper (ICASSP 2022, author-reported): 3.6032 DNSMOS on the Interspeech 2021 DNS Challenge blind test set, best among complex-domain models
- 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 — 7/10
Scored 7/10 (TIER metrics, metrics-v1). Evidence: peer-reviewed ICASSP 2022 paper reporting the best DNSMOS (3.6032) on the Interspeech 2021 DNS Challenge blind set — a result the challenge organizers measured, giving it more weight than pure self-evaluation. Strengths: joint enhancement and dereverberation, strong dual-domain design. Docked for: (1) the promised public code could not be confirmed at research time — no official weights; (2) non-causal, offline only. 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
- 5.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · ICASSP 2022 paper (arXiv 2111.06015); U-Net + dilated complex&real dual-path conformer
- · 3.6032 DNSMOS on Interspeech 2021 DNS Challenge blind test set — best among complex-domain models (challenge-measured, author-reported)
- · Promised public code could not be confirmed at research time; non-causal, offline only