code/README.md

# Code: extraction + analysis

This directory contains everything needed to (a) regenerate the pre-extracted embeddings from audio, and (b) reproduce the figures and tables in the accompanying NeurIPS 2026 paper.

## Layout

```
code/
  extraction_utils.py             # shared audio loading and save logic
  extract_*.py                    # 10 per-model extraction scripts
  extract_ssl_layers.py           # per-transformer-layer extraction (5 SSL models)
  run_all_extractions.sh          # master runner
  benchmark_analysis.ipynb        # main analysis notebook (80 cells)
  reproduce.sh                    # end-to-end reproduction (default: analysis only)
  README.md                       # this file
```

## Quick start

```bash
pip install -r ../requirements.txt
cd code
bash reproduce.sh
```

`reproduce.sh` defaults to **analysis-only**: it executes the notebook against the embeddings already shipped in `../data/embeddings/`. This takes ~10 minutes on a laptop.

To re-extract embeddings from the audio:

```bash
bash reproduce.sh --extract        # ~24 CPU-hours + ~1 GPU-hour
```

The notebook auto-resolves paths via `Path.cwd().parent`, so just open it from `code/` (`jupyter notebook benchmark_analysis.ipynb`) or run via the command above.

## Path resolution

All scripts and the notebook expect the release directory layout:

```
<VIPBENCH_ROOT>/
  code/                <-- you are here
  data/audio/reference/*.wav
  data/audio/comparison/*.wav
  data/embeddings/<model>.npz             <-- output of extraction
  data/embeddings/layers/<model>.npz      <-- per-layer SSL output
```

`extraction_utils._resolve_root()` picks the root via:

1. `VIPBENCH_ROOT` environment variable (if set), else
2. parent of the script's directory.

Override with `VIPBENCH_ROOT=/some/other/path bash reproduce.sh`.

## Models

| Model | Script | HF checkpoint | Dim | Type |
|---|---|---|---|---|
| RawNet3 | `extract_rawnet3_embeddings.py` | espnet/voxcelebs12_rawnet3 | 192 | Supervised |
| ECAPA-TDNN | `extract_ecapa_tdnn.py` | speechbrain/spkrec-ecapa-voxceleb | 192 | Supervised |
| TitaNet | `extract_titanet.py` | nvidia/speakerverification_en_titanet_large | 192 | Supervised |
| x-vector | `extract_xvector.py` | speechbrain/spkrec-xvect-voxceleb | 512 | Supervised |
| resemblyzer | `extract_resemblyzer.py` | (bundled with package) | 256 | Supervised |
| wav2vec 2.0 | `extract_wav2vec2.py` | facebook/wav2vec2-base | 768 | SSL |
| HuBERT | `extract_hubert.py` | facebook/hubert-base-ls960 | 768 | SSL |
| WavLM | `extract_wavlm.py` | microsoft/wavlm-base-plus | 768 | SSL |
| XLS-R | `extract_xlsr.py` | facebook/wav2vec2-xls-r-300m | 1024 | SSL |
| Whisper | `extract_whisper.py` | openai/whisper-base (encoder) | 512 | Weakly supervised |

Per-layer mean-pooled embeddings for the 5 SSL models are produced by `extract_ssl_layers.py` and saved to `data/embeddings/layers/<model>.npz`.

## Output format

Each `data/embeddings/<model>.npz` is a key-value store keyed by audio basename without `.wav` (e.g. `M01R`, `1_F01`). Values are 1-D `np.float32` arrays of shape `(embedding_dim,)`. The 9,900 keys cover 100 references plus 9,800 comparisons.

Per-layer bundles (`layers/<model>.npz`) use the same keys; values have shape `(num_layers, embedding_dim)`.

## Notes

- TitaNet requires NVIDIA NeMo (`nemo_toolkit[asr]`); install is heavy (~5 GB). The line is commented out in `requirements.txt`.
- The notebook caches expensive computations under `code/cache/`. Delete `code/cache/` to force recompute.
- Models are downloaded from Hugging Face on first run; subsequent runs use the local cache.

## License

Code in this directory is MIT-licensed (see `../LICENSE-CODE`). The dataset (audio, judgments, embeddings) is CC-BY-NC 4.0 (see `../LICENSE`).