TF-GridNet
The model that made time-frequency-domain separation great again — 23.4 dB SI-SDRi on WSJ0-2mix with a multi-path spectro-temporal grid.
About
By 2022 the field had mostly moved to time-domain models like SepFormer and assumed the spectrogram was dead. TF-GridNet proved otherwise. It stacks multi-path blocks where each block has three cooperating modules: an intra-frame spectral module that scans across frequencies within a frame, a sub-band temporal module that scans across time within each frequency band, and a full-band self-attention module for global context. Trained for complex spectral mapping with an SI-SDR loss plus a novel mixture-consistency term (the separated sources must add back up to the input), it hit 23.4 dB SI-SDR improvement on WSJ0-2mix without dynamic mixing — a large jump over the previous best. The extended TASLP 2023 version pushed further: state-of-the-art on SMS-WSJ and WHAMR! (noisy reverberant separation), multi-channel extensions, speech dereverberation, and the L3DAS22 enhancement challenge. The authors did not release official code, but the design is faithfully reimplemented in ESPnet and several community repos, and it became the backbone many later models (TF-Locoformer, TF-MossFormer) are measured against.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Paper (ICASSP 2023, author-reported): 23.4 dB SI-SDRi on WSJ0-2mix without dynamic mixing
- Paper (TASLP 2023, author-reported): SOTA on SMS-WSJ, WHAMR!, dereverberation and L3DAS22 enhancement tasks
- 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: ICASSP 2023 + TASLP 2023 papers from CMU with a landmark 23.4 dB SI-SDRi result; one of the most benchmarked-against architectures in separation since 2023. Strengths: broad SOTA coverage across separation, dereverberation, and multi-channel tasks; revived the T-F paradigm. Docked for: (1) no official code or weights from the authors — reproducibility depends on third-party ports with varying licenses; (2) every number is author-reported. License: none — no author release (community ports carry their own licenses).
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 8.0/10
- Maintenance · recency, releases, issue load
- 6.0/10
- Openness · license + docs
- 5.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · CMU / Hyundai 42dot; arXiv 2209.03952 (Sep 2022); ICASSP 2023 + TASLP 2023 extended version
- · 23.4 dB SI-SDRi on WSJ0-2mix without dynamic mixing (author-reported); mixture-consistency loss
- · No official author code — faithfully reimplemented in ESPnet and community repos; backbone for TF-Locoformer/TF-MossFormer