docs/model_table.md

# Baseline models

The 10 speech representations benchmarked in the accompanying paper. Each is loaded from a publicly available pretrained checkpoint; weights are downloaded by the corresponding extraction script in `code/`.

| Model | Paradigm | Training data | Dim | HF checkpoint | Extraction script |
|---|---|---|---|---|---|
| x-vector | Supervised classification | VoxCeleb 1+2 | 512 | `speechbrain/spkrec-xvect-voxceleb` | `extract_xvector.py` |
| ECAPA-TDNN | AAM-Softmax | VoxCeleb 1+2 | 192 | `speechbrain/spkrec-ecapa-voxceleb` | `extract_ecapa_tdnn.py` |
| RawNet3 | AAM-Softmax | VoxCeleb 1+2 | 192 | `espnet/voxcelebs12_rawnet3` | `extract_rawnet3_embeddings.py` |
| TitaNet (large) | AAM-Softmax | VoxCeleb 1+2, Fisher, SWB, LibriSpeech, NIST SRE | 192 | `nvidia/speakerverification_en_titanet_large` | `extract_titanet.py` |
| resemblyzer | GE2E loss, 3-layer LSTM | LibriSpeech + VoxCeleb 1+2 | 256 | bundled with `resemblyzer` package | `extract_resemblyzer.py` |
| wav2vec 2.0 | Contrastive masked prediction | LibriSpeech 960 h | 768 | `facebook/wav2vec2-base` | `extract_wav2vec2.py` |
| HuBERT | Masked prediction | LibriSpeech 960 h | 768 | `facebook/hubert-base-ls960` | `extract_hubert.py` |
| WavLM | Masked prediction + denoising | 94K h mixed | 768 | `microsoft/wavlm-base-plus` | `extract_wavlm.py` |
| XLS-R | Contrastive multilingual | 436K h, 128 languages | 1024 | `facebook/wav2vec2-xls-r-300m` | `extract_xlsr.py` |
| Whisper (encoder, base) | Multitask weakly supervised ASR | 680K h web audio | 512 | `openai/whisper-base` | `extract_whisper.py` |

## Output format

Each script saves `<model>.npz` to `<VIPBENCH_ROOT>/data/embeddings/`. The file is a key-value store with 9,900 keys (audio basenames without `.wav`) mapping to 1-D `np.float32` arrays of shape `(embedding_dim,)`.

For SSL models (wav2vec 2.0, HuBERT, WavLM, XLS-R, Whisper), `extract_ssl_layers.py` additionally produces a per-layer mean-pooled bundle saved to `<VIPBENCH_ROOT>/data/embeddings/layers/<model>.npz`. Values there have shape `(num_layers, embedding_dim)`.

## Pooling

Self-supervised models (wav2vec 2.0, HuBERT, WavLM, XLS-R) and Whisper produce frame-level outputs; we use **mean pooling across time** for the utterance-level embedding. Speaker-specialized models (x-vector, ECAPA-TDNN, RawNet3, TitaNet, resemblyzer) produce a single utterance-level vector by design and are passed through unchanged.

## Best-layer protocol

For SSL models the final layer is known to underperform on speaker tasks (SUPERB, 2021). The notebook implements a **SUPERB-style nested speaker-CV best-layer protocol**: for each held-out speaker fold, the layer that maximizes Pearson r against `P(same)` on training speakers is selected and applied to the test speakers. No pair contributes to selecting its own layer. Per-layer Pearson r curves are reported in the appendix figure.

## Licenses

Each baseline retains its original license. As of release time:

- speechbrain checkpoints: Apache 2.0
- ESPnet checkpoints: Apache 2.0
- NVIDIA NeMo: NVIDIA Source Code License (research permitted)
- Hugging Face hosted facebook/microsoft checkpoints: respective lab licenses (typically MIT / Apache 2.0)
- OpenAI Whisper: MIT
- resemblyzer: MIT

Verify the license at the linked checkpoint page before redistribution. The CC-BY-NC 4.0 license on VIPBench's audio + judgments + derived embeddings does not extend to these third-party model weights.