D3Net
Sony's densely-connected multidilated DenseNet — the 2020 MUSDB18 benchmark that DenseNet-era separation is measured against.
About
D3Net is Sony's answer to a specific architectural tension: dilated convolutions give you a wide receptive field cheaply, but naive dilation creates aliasing artifacts; dense skip connections give you rich feature reuse, but they bloat the model. D3Net combines multidilated convolution with dense skip connections in a single layer, modeling local and global structure simultaneously, with an explicit fix for the dilation-aliasing problem.\n\nOn MUSDB18 the authors report 7.80 vocals / 7.36 drums / 6.20 bass / 5.37 other / 6.68 average SDR with 1,500 songs of extra training data — state-of-the-art for its moment, and a clear step past the MMDenseNet/LSTM hybrids Sony had fielded at SiSEC 2018.\n\nIt is an arXiv preprint (2020) rather than a peer-reviewed publication, and there is no confirmed official code or weight release from Sony, so treat it as a well-documented research design rather than a runnable artifact. Its value in a catalog is as the reference point for the DenseNet generation of separators.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Author-reported (arXiv 2010.01733, 2020): MUSDB18 SDR with extra data — vocals 7.80, drums 7.36, bass 6.20, other 5.37, avg 6.68 dB; 6.01 dB avg without extra data\n- 2020-era numbers; no independent replication published and no official weight release
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 verified, metrics-v1). Evidence: detailed arXiv paper (2010.01733) from Sony's audio team with full MUSDB18 ablations and a genuine architectural contribution (aliasing-aware multidilated DenseNet). Strengths: the reference DenseNet-era separator, clear SOTA-at-the-time claim with extra-data numbers. Docked for: (1) preprint only — no peer review; (2) no confirmed official code or weight release; (3) 2020-era and fully superseded. License: research publication (CC BY 4.0 on the paper); no software artifact.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 5.0/10
- Maintenance · recency, releases, issue load
- 3.0/10
- Openness · license + docs
- 6.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · arXiv preprint 2010.01733 (Oct 2020) from Naoya Takahashi and Yuki Mitsufuji (Sony): multidilated convolution plus dense skip connections in one layer, with an explicit fix for dilation aliasing.
- · Author-reported MUSDB18 SDR: vocals 7.80, drums 7.36, bass 6.20, other 5.37, avg 6.68 with 1,500 songs of extra data (6.01 avg without).
- · Preprint only, no peer review, no official artifact, 2020-era and fully superseded — included as the reference DenseNet-era separator.