nsnet_dse — run artefacts
Reproduction artefacts for the nsnet_dse toolchain: four int8 /
DRUM speech-enhancement engines (NSNet2 dense, Monarch sparse NSNet2, ConvFSENet, Gated
ConvFSENet) and the DRUM approximate-multiplier design-space exploration built on them.
The code repo is self-contained except for these artefacts (trained checkpoints,
result outputs, and a bit-exact engine snapshot), which are gitignored. The repo's
common/hf_artifacts.ensure_artifact() pulls the checkpoints from here automatically on a
fresh clone, so git clone → reproduce works with no manual downloads.
Dataset (not hosted here): JacobLinCool/VoiceBank-DEMAND-16k
— 11,572 train / 824 test utterances at 16 kHz.
Layout
checkpoints/ trained weights (the load-bearing artefacts)
nsnet2_rt_gain8_best.pt NSNet2 FP32 (input-gain-8 recipe)
convfsenet_qf_causal_best.pt ConvFSENet FP32 (also the gated warm-start)
convfsenet_gated_causal_best.pt Gated ConvFSENet FP32
monarch_full_g_best.pt Monarch FP32 source (eco8 cp_monarch_full/g_best)
monarch_full_qat.npz Monarch packed int8 QAT weights (the deployed model)
results/ per-utt CSVs, paired-CI txt, train metrics/logs, and the
sensitivity/defensible/clamp-decomposition JSONs + a repro log
engines/ bit-exact snapshot: per-model int8 header (*.h) + compiled *_bin
MANIFEST.json file -> {sha256, bytes, producer script, consumers}
Verified headline numbers (full 824-utt VoiceBank-DEMAND test, real C engine)
| model | int8 PESQ-WB | STOI | reference |
|---|---|---|---|
| NSNet2 (dense) | 2.8116 | 0.9354 | FP32 recipe, input-gain 8 |
| Monarch (sparse, QAT) | 2.7443 | 0.9309 | FP32 ref 2.8374 |
| ConvFSENet | 2.7556 | 0.9179 | FP32 ref 2.7672 |
| Gated ConvFSENet (50-utt) | 2.9047 | 0.9233 | FP32 ref 2.8852 |
(Reproduced 2026-07-10, full 824-utt test unless noted; matches docs/GATE-RESULTS.md.)
Numbers and the paired A/B comparisons are recorded in
results/REPRODUCE-<date>.md and cross-checked against docs/GATE-RESULTS.md in the code
repo. Energy percentages in the code repo are multiplier-array-only with unverified
per-op constants — read them as a relative ranking, not system energy.
Reproduce
git clone <code-repo> && cd nsnet_dse
pip install -r requirements.txt
# checkpoints auto-download from this repo on first use, or grab them all:
hf download claroche1/nsnet-dse-artifacts --local-dir artifacts
cd nsnet2 && bash build.sh && python3 reproduce_eval.py # etc. per model
Provenance caveats
Do not mix checkpoints / engines / pipeline vintages — see docs/RESULTS-PROVENANCE.md
in the code repo. monarch_full_qat.npz is the authoritative deployed Monarch model;
monarch_full_g_best.pt is its FP32 source. *_last.pt optimizer-state snapshots and the
unreferenced convfsenet_snapshot.pt are intentionally omitted.