PURL for Noisy Correspondence SV (Pre-release)

This repository contains code for:

"PURL: Pairwise Unlearning with Reliability Learning for Noisy Correspondence in Supervised Contrastive Speaker Verification"
Anonymous submission to Interspeech 2026.

Project Status

This repository is a pre-release research codebase.

• It is being prepared for public release and reproducibility.
• The corresponding Interspeech 2026 submission is under review (not yet accepted).
• APIs, configs, and default paths may still change.

Method Overview (Paper-aligned)

PURL is a post-hoc, pair-level unlearning framework for supervised contrastive speaker verification under noisy correspondence.

• Problem: Positive pairs can be mismatched in large-scale data, which corrupts pairwise supervision.
• Stage 1 (reliability estimation): Fit a 2-component GMM on pairwise cosine similarities in a pretrained embedding space and compute pair confidence w_ij.
• Stage 2 (uncertainty-guided update):
  • retain reliable pairs with SupCon objective
  • suppress unreliable pairs with weighted cosine repulsion proportional to (1 - w_ij)
• Goal: reduce harmful correspondences while preserving clean intra-speaker compactness.

In this codebase, this is implemented through NCUTrainer with ncu_loss_type: soft.

What Is Included

This public pre-release keeps four training configs:

• configs/supcon.yaml
• configs/supcon_resnet34.yaml
• configs/ncu_30ep.yaml
• configs/ncu_30ep_resnet34.yaml

All machine-specific absolute paths were replaced with placeholders. Update them before running.

Experiments in This Release

The retained configs correspond to the two backbones and post-hoc setup used in the paper:

• ECAPA-TDNN
  • base SupCon pretraining: configs/supcon.yaml
  • 30-epoch post-hoc update (PURL/NCU): configs/ncu_30ep.yaml
• Thin ResNet-34
  • base SupCon pretraining: configs/supcon_resnet34.yaml
  • 30-epoch post-hoc update (PURL/NCU): configs/ncu_30ep_resnet34.yaml

Environment Setup

Python 3.10 is recommended.

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
pip install -e .

If you use conda:

conda create -n chns python=3.10 -y
conda activate chns
pip install -r requirements.txt
pip install -e .

Important Runtime Notes

1. Configs use class paths like trainers.*, so run with:

PYTHONPATH=src python run_train.py ...
PYTHONPATH=src python run_test.py ...

2. NCU/PURL post-hoc runs assume a pretrained SupCon checkpoint via --ckpt_path.

Dataset Preparation

1. Download VoxCeleb data from the official source: VoxCeleb download page.
2. Convert source audio to the format expected by your experiments (commonly 16 kHz mono wav/flac).
3. Create spk2utt mapping files:

./scripts/make_spk2utt_and_utt2spk.sh /path/to/vox2_dev_wav/wav

4. Edit each config to set placeholders such as:

• /path/to/vox2_dev_wav/wav
• /path/to/vox2_dev_wav/wav/spk2utt
• /path/to/vox1_test_wav/wav
• /path/to/voxceleb1_H.txt

5. Ensure spk2utt files are consistent with your data layout and path conventions.

Train

SupCon (ECAPA):

PYTHONPATH=src python run_train.py --config configs/supcon.yaml

SupCon (ResNet34):

PYTHONPATH=src python run_train.py --config configs/supcon_resnet34.yaml

PURL post-hoc update (ECAPA, 30 epochs):

PYTHONPATH=src python run_train.py --config configs/ncu_30ep.yaml --ckpt_path /path/to/base_supcon.ckpt

PURL post-hoc update (ResNet34, 30 epochs):

PYTHONPATH=src python run_train.py --config configs/ncu_30ep_resnet34.yaml --ckpt_path /path/to/base_supcon_resnet34.ckpt

Evaluation

PYTHONPATH=src python run_test.py --config configs/ncu_30ep.yaml --ckpt_name epoch=XXX_step=YYYY

or

PYTHONPATH=src python run_test.py --config configs/supcon.yaml --ckpt_name epoch=XXX_step=YYYY

--ckpt_name should be the checkpoint filename stem under <save_dir>/checkpoints.

Reproducibility Notes

• run_train.py supports --ckpt_path for explicit resume.
• run_test.py expects --ckpt_name and uses the checkpoint directory from logger config.
• For smoke tests with limit_test_batches or fast_dev_run, only partial trial embeddings are computed.
• ncu_30ep*.yaml currently uses voxceleb1_H-style placeholder for default test trials.

Planned Updates

• Add final camera-ready paper link and BibTeX after decision/publication.
• Add exact benchmark scripts and standardized evaluation table templates for easier replication.

Citation

If you use this code, please cite the final paper once available.

Temporary reference:

PURL: Pairwise Unlearning with Reliability Learning for Noisy Correspondence in Supervised Contrastive Speaker Verification (Interspeech 2026 submission, under review)