HaoweiXu0126/SubCI-GTCRN

Data and Model Artifacts

The reproducibility assets for this repository are indexed in one Hugging Face repository:

• Hub repository: https://huggingface.co/HaoweiXu0126/SubCI-GTCRN
• File browser: https://huggingface.co/HaoweiXu0126/SubCI-GTCRN/tree/main

The Hub repository contains the compact project-owned artifacts required to inspect or reuse the reported model family. Public third-party corpora remain at their authoritative sources instead of being redistributed. TICI recordings are not uploaded because the repository does not contain an explicit public-redistribution license for those recordings.

Checkpoints

The official bundle is subci_physcue_official_20260503_112250_689114. The uploaded checkpoints cover all three reported seeds and keep the acoustic baseline, offline SubCI candidate, and distilled deployment student separate.

Role	Seed	Size (bytes)	SHA-256	Download
Acoustic baseline	43	436394	1255faefaccc7dba36c83b285c67c23fadcd833ca3bf76ef7b79144f442ba1d8	best.pt
Acoustic baseline	44	436394	7674915fbf884110a8a935b43684e1d6adaaff83bf2f1eba52aeb1d2140f8f32	best.pt
Acoustic baseline	45	436394	bd44752d323c8fa93d96d5e415d0461c106f61fabca8855a593b28c5bf4be9c5	best.pt
Offline SubCI candidate	43	2012241	2e15d2250b704fc13585409a1b59bbd7ce77f144b2b5ca04f5b9e42981799984	best.pt
Offline SubCI candidate	44	2012241	06484b6a0b32b9842256dd66b5227a0962deda989809fc5634aa52ccce8dd223	best.pt
Offline SubCI candidate	45	2012241	87fd53e6e8882e52218f3ae21bd5949da5e741e043aa49d7d3806a619c4e97dc	best.pt
Distilled deployment student	43	1380980	f695c08543e0cf408d5bc443d95899d04539c8a0421276eadb61a9ad56a1861c	best.pt
Distilled deployment student	44	1380980	6c0198f6e9ca734dfb85c9129684965c9ee2673aae89c0cc9676fb2171ff3cb0	best.pt
Distilled deployment student	45	1380980	b34f9db0cde14b55820ae1283a4564a20a4657deac4874b160af66dec57960d2	best.pt

The seed-43 deployment export is also available in inference formats:

Format	Size (bytes)	SHA-256	Download
TorchScript	562615	0a35678237f6149f81aa30b8a2de2a1b9f9a63ec88ff94b9cc7be2c8d97c318f	model.ts
ONNX	415691	1953417154e31c1f7f7f51744d473d3a3e5c3d44d6c8fe36f584d60037f785ac	model.onnx

These artifacts do not imply that the deployment gate passed. The governing official bundle reports relative_gate_ok=true, absolute_gate_ok=false, deploy_gate_ok=false, and final_pass=false.

Public Data Sources

Data source	Project use	Authoritative link
VoiceBank-DEMAND noisy speech database	Paired clean/noisy speech and public clean-speech source	https://datashare.ed.ac.uk/handle/10283/2791
CSTR VCTK Corpus	Clean speech used by the VCTK+TICI workflow	https://datashare.ed.ac.uk/handle/10283/3443
Microsoft DNS Challenge	Public speech, noise, and impulse-response material	https://github.com/microsoft/DNS-Challenge
OpenSLR RIR and Noise Database	Public room impulse responses and noises	https://www.openslr.org/28/

The exact public downloads used by the local preparation workflow are declared in:

• SEtrain/configs/tici_optimization/public_audio_download_bootstrap.yaml
• SEtrain/configs/tici_optimization/public_audio_download_dns_noise_ir_bootstrap.yaml
• SEtrain/configs/tici_optimization/public_data_manifest.yaml

TICI Data Boundary

The local data/TICI_data/ directory contains body-noise recordings used by the VCTK+TICI stress workflow. These recordings are intentionally absent from the public Hub repository because no explicit redistribution license is recorded in this codebase. The repository publishes the preparation and evaluation code, but access to the underlying TICI recordings must follow the data owner's approved access process.

Command-Line Download

Download the complete compact artifact repository:

hf download HaoweiXu0126/SubCI-GTCRN \
  --local-dir artifacts/SubCI-GTCRN

Download only the primary offline checkpoint:

hf download HaoweiXu0126/SubCI-GTCRN \
  checkpoints/offline_subci/seed_43/best.pt \
  --local-dir artifacts/SubCI-GTCRN

After download, verify a file with:

sha256sum artifacts/SubCI-GTCRN/checkpoints/offline_subci/seed_43/best.pt