MossFormer2

MossFormer2 is Alibaba SpeechLab's answer to a real weakness in attention-heavy separators: self-attention is great at long-range coarse structure but sloppy at fine-grained recurrent patterns. The fix is a hybrid — the MossFormer module (full attention over local segments plus a linearized efficient attention over the whole sequence) paired with a novel recurrent module built not from RNNs but from feed-forward sequential memory networks (FSMN) with gated convolutional units and dense connections, so it captures recurrence without recurrent connections and trains in parallel. The authors report 24.1 dB SI-SDRi on WSJ0-2mix with dynamic mixing, plus wins on WSJ0-3mix, Libri2Mix, and WHAM!/WHAMR! — the large config is 55.7M parameters at a reported 0.053 real-time factor. Unlike many paper models, this one is genuinely usable: pretrained checkpoints ship through ClearerVoice-Studio (Apache-2.0) with enhancement, separation, and super-resolution variants on Hugging Face, including a 48 kHz full-band speech enhancement model. It also landed on ModelScope with millions of reported uses. That deployment story is why it scores above most research-only separators.

• Paper (arXiv 2312.11825, author-reported): 24.1 dB SI-SDRi on WSJ0-2mix (large, with dynamic mixing), strong results on WSJ0-3mix, Libri2Mix, WHAM!/WHAMR!
• No independent benchmark replication found