Add files using upload-large-folder tool

CHANGELOG.md

@@ -0,0 +1,14 @@
+# Changelog
+
+## v1.0 (2026-05-06)
+
+Initial release for NeurIPS 2026 Evaluations & Datasets Track submission.
+
+- 124,876 listener judgments from 1,290 English-speaking participants on 9,800 voice pairs across 100 speakers.
+- 9,900 audio files: 100 reference clips + 9,800 comparison clips (16 kHz mono WAV).
+- Pre-extracted embeddings for 10 speech models: x-vector, ECAPA-TDNN, RawNet3, TitaNet, resemblyzer (supervised); wav2vec 2.0, HuBERT, WavLM, XLS-R (self-supervised); Whisper (weakly supervised).
+- Per-layer mean-pooled embeddings for the 5 SSL models.
+- Croissant 1.0 metadata with both core and Responsible AI fields.
+- Datasheet for Datasets, annotation protocol, and reproduction documentation.
+
+Anonymized for double-blind review. Author identities and DOI to be added at camera-ready.

CITATION.cff

@@ -0,0 +1,27 @@
+cff-version: 1.2.0
+message: "If you use VIPBench in your research, please cite the accompanying NeurIPS 2026 paper."
+title: "VIPBench: A Human-Aligned Benchmark for Voice Identity Perception in the Age of Voice Cloning"
+abstract: "VIPBench is a benchmark of 124,876 same/different identity judgments from 1,290 English-speaking listeners on 9,800 voice pairs across 100 speakers, spanning real recordings, AI voice clones, and continuously morphed voices. The release includes audio, listener judgments, and pre-extracted embeddings for ten speaker and speech-representation models. Four evaluation tasks measure model alignment with human voice-identity perception."
+type: dataset
+authors:
+  - family-names: "Anonymous"
+    given-names: "Authors"
+    affiliation: "Withheld for double-blind review"
+license: CC-BY-NC-4.0
+version: "1.0"
+date-released: "2026-05-06"
+keywords:
+  - speaker embeddings
+  - voice identity perception
+  - human-aligned benchmark
+  - voice cloning
+  - perceptual evaluation
+preferred-citation:
+  type: conference-paper
+  title: "VIPBench: A Human-Aligned Benchmark for Voice Identity Perception in the Age of Voice Cloning"
+  authors:
+    - family-names: "Anonymous"
+      given-names: "Authors"
+  collection-title: "Advances in Neural Information Processing Systems Datasets and Benchmarks (NeurIPS Evaluations and Datasets Track)"
+  year: 2026
+  notes: "Anonymized for double-blind review. Author identities and DOI to be added at camera-ready."

DATASHEET.md

@@ -0,0 +1,142 @@
+# Datasheet for VIPBench
+
+Following the structure of Gebru et al. 2021, "Datasheets for Datasets" (Communications of the ACM 64(12)).
+
+## Motivation
+
+**For what purpose was the dataset created?**
+To benchmark whether speaker-embedding and speech-representation models align with human voice-identity perception. Existing speaker benchmarks evaluate against a metadata speaker label (who produced the recording); this dataset measures whether listeners hear two voices as the same speaker, which can diverge from the metadata label especially under voice cloning and voice morphing.
+
+**Who created the dataset and on behalf of which entity?**
+Withheld for double-blind review. Will be filled in at camera-ready.
+
+**Who funded the creation of the dataset?**
+Withheld for double-blind review.
+
+## Composition
+
+**What do the instances represent?**
+Each instance is a voice pair (one reference audio clip plus one comparison audio clip) annotated with same-or-different-speaker judgments collected from human listeners.
+
+**How many instances are there?**
+9,800 voice pairs, 124,876 listener judgments, 1,290 listeners, 100 speakers, 9,900 audio files (100 reference + 9,800 comparison).
+
+**Does the dataset contain all possible instances or is it a sample?**
+It is a designed sample. 100 speakers were selected to balance 5 sociophonetic groups x 2 genders x 2 age brackets, 5 speakers per cell. Pairs were sampled to cover six stimulus types: same recording, same speaker different recording, same speaker AI clone, different speaker different recording, different speaker AI clone, and continuously morphed pairs.
+
+**What data does each instance consist of?**
+- A reference audio clip (16 kHz mono WAV, ~6 seconds).
+- A comparison audio clip (16 kHz mono WAV, similar duration).
+- Per-pair aggregates (`stimuli.csv`): stimulus type, source speaker(s), morph scale (Type 6 only), counts of same/different votes, accuracy against the metadata label, derived `P(same)`.
+- Per-judgment records (`participant_responses.csv`): listener ID, stimulus ID, binary same/different answer, listener-recognition flag, listener demographics (age band, gender, first language).
+
+**Is there a label or target associated with each instance?**
+The primary target is the human consensus `P(same)` per pair. A binary majority-vote label (`y = 1[P(same) > 0.5]`) is also supplied. The ground-truth metadata speaker label is included for reference.
+
+**Is any information missing from individual instances?**
+Listener demographics include age band (categorical), gender, and first-language flag. Detailed listener demographics or browser/device information are not included. The `know_speaker` field is missing for some early-trial responses; trials with the listener flagging recognition are still released and noted.
+
+**Are relationships between individual instances made explicit?**
+Yes. `stimuli.csv` indexes pairs by reference and comparison speaker IDs; `participant_responses.csv` indexes by listener ID and stimulus ID. The 9,900-key embedding files use the audio basename as key, allowing exact join with `stimuli.csv.id`.
+
+**Are there recommended data splits?**
+Speaker-level cross-validation (10 folds, gender-balanced 5M+5F per fold) is recommended to avoid speaker leakage. The benchmark protocol uses speaker-level GroupKFold throughout.
+
+**Are there any errors, sources of noise, or redundancies?**
+Identity perception is intrinsically noisy: split-half Spearman-Brown reliability of `P(same)` is rho_SB = 0.705, bounding any model's correlation against the observed target. Voice clones and mid-morph stimuli are designed to be ambiguous, so consensus is itself probabilistic.
+
+**Is the dataset self-contained or does it link to external resources?**
+Self-contained for evaluation: audio, judgments, and pre-extracted embeddings are bundled. Pretrained model weights are downloaded from Hugging Face / model hubs by the extraction scripts; their checkpoint identifiers are fixed in `docs/model_table.md`.
+
+**Does the dataset contain confidential data?**
+No. Listener identifiers are pseudonymized integers tied to no external account. Speaker names are public figures whose recordings are publicly available.
+
+**Does the dataset contain offensive, insulting, or threatening content?**
+No. Stimuli are conversational speech.
+
+## Collection process
+
+**How was the data acquired?**
+- Reference audio: selected from publicly available recordings (interviews, podcasts) of 100 US celebrities.
+- Comparison audio: a mix of additional clips of the same speakers, AI voice clones generated with a state-of-the-art TTS system from a short reference clip per speaker, and continuously morphed pairs sweeping a voice-conversion latent.
+- Judgments: 1,290 adult English-speaking listeners recruited via the Centaur AI platform. Each pair received at least 8 judgments.
+
+**What mechanisms or procedures were used to collect the data?**
+Listeners heard each pair as a single audio clip (reference + 1 second silence + short beep + comparison) and answered (a) whether the two clips came from the same speaker (binary) and (b) optionally identified which of four within-group celebrities or "I don't know" they recognized in the reference.
+
+**Who was involved in the data collection process?**
+Listeners were paid via the Centaur AI platform per the platform's standard rate, which meets minimum-wage requirements in the country of data collection. The annotation protocol was approved by an Institutional Review Board.
+
+**Over what time frame was the data collected?**
+Stimulus creation and listening study were conducted between 2025 and 2026.
+
+**Were any ethical review processes conducted?**
+Yes. Both the clone generation (from publicly available celebrity speech) and the human-annotation study were covered under IRB review. Participant consent followed the platform's standard consent pipeline.
+
+## Preprocessing, cleaning, labeling
+
+**Was preprocessing or cleaning done?**
+- Audio was resampled to 16 kHz mono.
+- Listener responses were checked against attention-check probes; per-listener qualification flags are included in `participant_responses.csv`.
+- The `reorganized_stimuli.csv` aggregates count same/different votes per pair.
+
+**Is the raw data saved?**
+Pre-resampling source audio is not included in the release; the IRB-covered raw participant responses are included as `participant_responses.csv`.
+
+## Uses
+
+**Has the dataset been used for any tasks already?**
+Yes. The accompanying NeurIPS 2026 paper benchmarks 10 publicly available speaker and speech-representation models against the four evaluation tasks (continuous `P(same)` prediction, binary verification, RSA, real-to-synthetic transfer).
+
+**Is there a repository linking to papers using the dataset?**
+Will be maintained at the public repository to be created at camera-ready.
+
+**What other tasks could the dataset be used for?**
+- Listener-conditioned identity modeling.
+- Voice-clone perceptual evaluation for TTS development (under the non-commercial license).
+- Calibration and uncertainty estimation in speaker verification.
+- Cross-language perception research, using these stimuli as a comparison anchor.
+
+**Is there anything that should not be done with the dataset?**
+- Do not use the audio, judgments, or derived embeddings for commercial purposes (CC-BY-NC 4.0).
+- Do not extract clone-generation recipes from the released audio. The benchmark intentionally does not release voice-conversion or TTS training targets.
+- Do not deanonymize listeners.
+
+## Distribution
+
+**Will the dataset be distributed to third parties?**
+Yes, openly via Hugging Face under CC-BY-NC 4.0.
+
+**How will it be distributed?**
+Hugging Face dataset (anonymized URL during review, permanent URL at camera-ready). Croissant metadata accompanies the release.
+
+**When will it be distributed?**
+Initial public release: 2026-05-06 (NeurIPS submission).
+
+**Will the dataset be distributed under a copyright or other intellectual property license?**
+- Audio, judgments, and embeddings: CC-BY-NC 4.0.
+- Code: MIT.
+- Pretrained model weights retain their original licenses.
+
+**Have any third parties imposed restrictions on the data?**
+No.
+
+## Maintenance
+
+**Who will maintain the dataset?**
+The authors. Contact information will be filled in at camera-ready.
+
+**How can the maintainers be contacted?**
+A mailing address will be provided at camera-ready. During review, author identification is withheld.
+
+**Is there an erratum?**
+Not yet. Errata will be tracked in `CHANGELOG.md`.
+
+**Will the dataset be updated?**
+Yes. Updates will be versioned. Older versions will remain accessible.
+
+**Will older versions of the dataset continue to be supported?**
+Yes, by maintaining tagged releases on the public repository.
+
+**If others want to extend the dataset, is there a mechanism for them to do so?**
+Yes. Researchers can collect additional listener judgments using the protocol documented in `docs/annotation_protocol.md` and contribute via the public repository (instructions to be added at camera-ready).

LICENSE

@@ -0,0 +1,19 @@
+VIPBench dataset is licensed under the Creative Commons Attribution-NonCommercial 4.0 International License (CC-BY-NC 4.0).
+
+Full license text: https://creativecommons.org/licenses/by-nc/4.0/legalcode
+Human-readable summary: https://creativecommons.org/licenses/by-nc/4.0/
+
+Under this license you are free to:
+  Share: copy and redistribute the material in any medium or format.
+  Adapt: remix, transform, and build upon the material.
+
+Under the following terms:
+  Attribution: You must give appropriate credit, provide a link to the license, and indicate if changes were made.
+  NonCommercial: You may not use the material for commercial purposes.
+  No additional restrictions: You may not apply legal terms or technological measures that legally restrict others from doing anything the license permits.
+
+This license applies to the audio files (reference, comparison, voice clones, voice morphs), human listener judgments, per-pair aggregates, speaker metadata, and pre-extracted embeddings distributed in this repository.
+
+Pretrained model weights loaded by the extraction scripts in code/ retain their original licenses; see docs/model_table.md.
+
+The code in code/ is licensed under the MIT License; see LICENSE-CODE.

LICENSE-CODE

@@ -0,0 +1,11 @@
+MIT License
+
+Copyright (c) 2026 VIPBench contributors (anonymized for NeurIPS 2026 double-blind review)
+
+Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+This license applies only to the source code in code/. The dataset (audio, judgments, embeddings, metadata) is licensed separately under CC-BY-NC 4.0; see LICENSE.

README.md

@@ -0,0 +1,181 @@
+---
+license: cc-by-nc-4.0
+language:
+- en
+pretty_name: VIPBench
+size_categories:
+- 100K<n<1M
+tags:
+- speaker-recognition
+- voice-identity
+- voice-cloning
+- human-perception
+- benchmark
+- audio
+task_categories:
+- audio-classification
+- other
+---
+
+# VIPBench: A Human-Aligned Benchmark for Voice Identity Perception in the Age of Voice Cloning
+
+VIPBench is a benchmark of **124,876 same/different identity judgments** from **1,290 English-speaking listeners** on **9,800 voice pairs** spanning **100 demographically-stratified speakers**. Pairs cover three stimulus families: real recordings, AI voice clones generated by a state-of-the-art TTS system, and continuously morphed voices.
+
+The benchmark evaluates whether speaker-embedding and speech-representation models align with human voice-identity perception, providing a perceptual evaluation target distinct from metadata-label speaker identification.
+
+> **Anonymized release for NeurIPS 2026 Evaluations & Datasets Track double-blind review.**
+> Author identities and permanent URLs will be added at camera-ready.
+
+---
+
+## Dataset summary
+
+| Item | Count |
+|---|---|
+| Speakers | 100 (50 M / 50 F, 5 sociophonetic groups, 2 age brackets) |
+| Reference audio clips | 100 (one per speaker) |
+| Comparison audio clips | 9,800 (98 per speaker) |
+| Voice pairs | 9,800 |
+| Listener judgments | 124,876 |
+| Listeners | 1,290 |
+| Median judgments per pair | 10 (range 8 to 92) |
+| Stimulus types | 6 (real same/different, AI clones, voice morphs) |
+| Pre-extracted speaker embeddings | 10 models |
+| Per-layer SSL embeddings | 5 models |
+
+## Supported tasks
+
+The benchmark defines four evaluation tasks:
+
+1. **Predict listener agreement rate** (continuous regression). Predict `P(same)` per pair. Metric: Pearson r, R^2 against the human consensus, bounded by the Spearman-Brown noise ceiling rho_SB = 0.705.
+2. **Human-aligned binary verification**. Classify pairs against the human majority vote. Metrics: AUC (ranking) and Platt-calibrated ECE (calibration).
+3. **Representational similarity (RSA)**. Spearman correlation between human and model representational dissimilarity matrices, with a Mantel permutation test.
+4. **Real-to-synthetic transfer**. Whether a predictor fit on real-speech pairs still works on voice clones and morphs.
+
+A 10-fold gender-balanced speaker-level cross-validation protocol prevents speaker leakage.
+
+## Dataset structure
+
+```
+data/
+  speakers.csv                   # 100 rows: speaker id, name, group, gender, age
+  stimuli.csv                    # 9,800 rows: per-pair aggregates (P(same), votes, type)
+  participant_responses.csv      # 124,876 rows: per-judgment records
+  stimuli_interpol.csv           # 8,100 rows: morph-trajectory metadata for Type 6
+  audio/
+    reference/                   # 100 *R.wav (16 kHz mono)
+    comparison/                  # 9,800 *.wav
+  embeddings/
+    rawnet3.npz, ecapa_tdnn.npz, titanet.npz, xvector.npz, resemblyzer.npz,
+    wav2vec2.npz, hubert.npz, wavlm.npz, xlsr.npz, whisper.npz
+    layers/                      # per-layer (mean-pooled) for SSL models
+      wav2vec2.npz, hubert.npz, wavlm.npz, xlsr.npz, whisper.npz
+samples/                         # 5-speaker quick-look subset (~150 MB)
+code/                            # 10 extraction scripts + analysis notebook + reproduce.sh
+docs/                            # annotation protocol, schemas, model table, reproduction
+```
+
+For column-level dictionaries see `docs/data_dictionary.md`. For the six stimulus types see `docs/stimulus_types.md`. For the listening-study protocol see `docs/annotation_protocol.md`.
+
+### Embedding format
+
+Each `.npz` is a key-value store keyed by audio basename without the `.wav` extension (e.g., `M01R`, `1_F01`, `4_M12_M15B`). Values are numpy arrays of shape `(embedding_dim,)` for the 10 main embeddings and `(num_layers, embedding_dim)` for the per-layer bundles. The 9,900 keys cover 100 references plus 9,800 comparisons.
+
+### Pairing reference and comparison
+
+Each row of `data/stimuli.csv` represents one voice pair. The `reference` column gives the reference speaker ID (e.g., `M01`) and the `id` column gives the stimulus identifier of the comparison clip (e.g., `1_M01`, `4_M12_M15B`). The pairing rule is:
+
+| You want | Reference clip | Comparison clip |
+|---|---|---|
+| Audio file | `data/audio/reference/{row.reference}R.wav` | `data/audio/comparison/{row.id}.wav` |
+| Embedding key | `{row.reference}R` (e.g., `M01R`) | `{row.id}` (e.g., `1_M01`) |
+
+Cosine-similarity scoring against `P(same)`:
+
+```python
+import numpy as np, pandas as pd
+from sklearn.metrics.pairwise import cosine_similarity
+
+stim = pd.read_csv('data/stimuli.csv')
+emb  = dict(np.load('data/embeddings/ecapa_tdnn.npz'))
+
+stim['cos'] = stim.apply(
+    lambda r: cosine_similarity(
+        emb[f'{r.reference}R'].reshape(1, -1),
+        emb[r.id].reshape(1, -1)
+    )[0, 0],
+    axis=1,
+)
+stim['p_same'] = stim['same_vote'] / stim['num_response']
+print(stim[['cos', 'p_same']].corr())   # Pearson r against listener consensus
+```
+
+## Quick start
+
+```bash
+pip install -r requirements.txt
+cd code && bash reproduce.sh        # ~10 min from cached embeddings
+```
+
+To re-extract embeddings from the audio (~24 CPU-hours plus ~1 GPU-hour for Whisper), see `docs/reproduction.md`.
+
+## Source data and collection
+
+- **Speakers.** 100 English-speaking US celebrities stratified across 5 sociophonetic groups (Italian-American, African-American, Asian-American, Latinx, mainstream Standard American English) x 2 genders x 2 age brackets, 5 speakers per cell.
+- **Reference audio.** Clips selected from publicly available recordings (interviews, podcasts).
+- **Voice clones.** Generated with a state-of-the-art TTS system (Cartesia Sonic) from a short reference clip per speaker.
+- **Voice morphs.** Continuously morphed pairs sweeping a voice-conversion latent between two source speakers, sampled at 11 levels per pair.
+- **Listeners.** 1,290 adult English-speaking participants recruited via the Centaur AI platform under an IRB-approved protocol. Consent followed the platform's standard pipeline.
+
+Each pair received at least 8 judgments; real-speech pairs (Types 1, 2, 4) received more coverage than synthetic pairs to give tighter consensus estimates on the real-speech reference distribution.
+
+## Considerations for use
+
+### Personally identifying information
+
+The dataset names public-figure speakers because the celebrity-stratified design is integral to the benchmark and source recordings are already public. Listener identifiers in `participant_responses.csv` are pseudonymized integers tied to no external account.
+
+### Biases and limitations
+
+- English-speaking listener pool, US-dialect speakers. Cross-language perception is not measured.
+- 100 speakers limits statistical power for some subgroup contrasts (20 speakers per sociophonetic group).
+- Studio-quality audio. In-the-wild conditions (noise, codec compression, telephony) are not represented.
+- The operational target is a population consensus, appropriate for ambiguous stimuli where any absolute identity label would itself be probabilistic.
+
+### Responsible use
+
+The benchmark measures model-human alignment at the evaluation level. We do not release clone-generation recipes or adversarial training targets. Voice-cloning systems that better align with human perception could inform adversarial use; the same alignment knowledge also strengthens defenses (perception-aligned identity models can flag clones that metadata-based verification accepts).
+
+## License
+
+- **Dataset (audio, judgments, metadata, embeddings):** Creative Commons Attribution-NonCommercial 4.0 International (CC-BY-NC 4.0). See `LICENSE`.
+- **Code (scripts, notebook):** MIT License. See `LICENSE-CODE`.
+- **Pretrained model weights** (loaded by extraction scripts): each baseline retains its original license; see `docs/model_table.md`.
+
+Commercial use of the audio, judgments, or derived embeddings is not permitted under this license.
+
+## Citation
+
+To be filled in at camera-ready.
+
+```
+@inproceedings{vipbench2026,
+  title  = {VIPBench: A Human-Aligned Benchmark for Voice Identity Perception in the Age of Voice Cloning},
+  author = {Anonymous},
+  booktitle = {Advances in Neural Information Processing Systems Datasets and Benchmarks},
+  year   = {2026},
+  note   = {Anonymized for double-blind review.}
+}
+```
+
+## Files
+
+- `README.md` (this file): dataset card.
+- `LICENSE`: CC-BY-NC 4.0 full text.
+- `LICENSE-CODE`: MIT full text for scripts.
+- `croissant.json`: MLCommons Croissant 1.0 metadata (core + Responsible AI fields).
+- `DATASHEET.md`: Datasheet for Datasets (Gebru et al. 2021).
+- `CHANGELOG.md`: version history.
+- `CITATION.cff`: machine-readable citation.
+- `requirements.txt`: pinned Python dependencies.
+- `data/`, `samples/`, `code/`, `docs/`: see structure section above.

code/README.md

@@ -0,0 +1,87 @@
+# 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`).

code/extract_ecapa_tdnn.py

@@ -0,0 +1,39 @@
+#!/usr/bin/env python3
+"""Extract ECAPA-TDNN embeddings using SpeechBrain.
+
+Model: speechbrain/spkrec-ecapa-voxceleb (supervised, AAM-Softmax, 192-dim)
+Install: pip install speechbrain
+"""
+
+import argparse
+import torch
+import numpy as np
+from extraction_utils import load_audio, extract_all
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--device", default="cuda" if torch.cuda.is_available() else "cpu")
+    parser.add_argument("--base-dir", default=None)
+    parser.add_argument("--output-dir", default=None)
+    args = parser.parse_args()
+
+    from speechbrain.inference.speaker import EncoderClassifier
+
+    print(f"Loading ECAPA-TDNN on {args.device}...")
+    classifier = EncoderClassifier.from_hparams(
+        source="speechbrain/spkrec-ecapa-voxceleb",
+        run_opts={"device": args.device},
+    )
+
+    def model_fn(audio_path):
+        audio = load_audio(audio_path, target_sr=16000)
+        signal = torch.tensor(audio).unsqueeze(0).to(args.device)
+        embedding = classifier.encode_batch(signal)
+        return embedding.squeeze().cpu().numpy()
+
+    extract_all(model_fn, "ecapa_tdnn", args.base_dir, args.output_dir)
+
+
+if __name__ == "__main__":
+    main()

code/extract_hubert.py

@@ -0,0 +1,43 @@
+#!/usr/bin/env python3
+"""Extract HuBERT embeddings using HuggingFace Transformers.
+
+Model: facebook/hubert-base-ls960 (self-supervised, 768-dim)
+Frame-level output is mean-pooled to get utterance-level embeddings.
+Install: pip install transformers
+"""
+
+import argparse
+import torch
+import numpy as np
+from extraction_utils import load_audio, extract_all
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--device", default="cuda" if torch.cuda.is_available() else "cpu")
+    parser.add_argument("--base-dir", default=None)
+    parser.add_argument("--output-dir", default=None)
+    args = parser.parse_args()
+
+    from transformers import HubertModel, Wav2Vec2FeatureExtractor
+
+    print(f"Loading HuBERT-base on {args.device}...")
+    feature_extractor = Wav2Vec2FeatureExtractor.from_pretrained("facebook/hubert-base-ls960")
+    model = HubertModel.from_pretrained("facebook/hubert-base-ls960").to(args.device)
+    model.eval()
+
+    def model_fn(audio_path):
+        audio = load_audio(audio_path, target_sr=16000)
+        inputs = feature_extractor(audio, sampling_rate=16000, return_tensors="pt")
+        inputs = {k: v.to(args.device) for k, v in inputs.items()}
+        with torch.no_grad():
+            outputs = model(**inputs)
+        hidden = outputs.last_hidden_state  # (1, T, 768)
+        embedding = hidden.mean(dim=1).squeeze().cpu().numpy()
+        return embedding
+
+    extract_all(model_fn, "hubert", args.base_dir, args.output_dir)
+
+
+if __name__ == "__main__":
+    main()

code/extract_rawnet3_embeddings.py

@@ -0,0 +1,242 @@
+#!/usr/bin/env python3
+"""
+Extract RawNet3 embeddings for all audio files.
+
+This script extracts RawNet3 speaker embeddings for:
+- 100 audio files ending with "R" in exp_2 directory (e.g., F01R.wav)
+- 9,800 audio files in output directory
+
+The embeddings are saved in a dictionary format where:
+- Key: filename without extension (e.g., "F01R")
+- Value: RawNet3 embedding vector (numpy array)
+
+Saved as both pickle (.pkl) and numpy (.npz) formats for flexibility.
+
+Requirements:
+- ESPnet-SPK: pip install espnet
+- Or RawNet repository: git clone https://github.com/jungjee/RawNet.git
+"""
+
+import os
+import sys
+import numpy as np
+from pathlib import Path
+import pickle
+import warnings
+
+warnings.filterwarnings('ignore')
+
+# Try to import tqdm for progress bars, fallback to basic iteration if not available
+try:
+    from tqdm import tqdm
+except ImportError:
+    def tqdm(iterable, desc=None):
+        if desc:
+            print(f"\nProcessing {desc}...")
+        return iterable
+
+# Try to import ESPnet first (preferred method)
+try:
+    from espnet2.bin.spk_inference import Speech2Embedding
+    ESPNET_AVAILABLE = True
+    print("ESPnet-SPK is available")
+except ImportError:
+    ESPNET_AVAILABLE = False
+    print("ESPnet-SPK not available. Please install it:")
+    print("  pip install espnet")
+    print("Or set up the RawNet repository manually.")
+    sys.exit(1)
+
+
+def load_audio(file_path, target_sr=16000):
+    """Load audio file and resample to target sample rate if needed."""
+    try:
+        # Try librosa first (most reliable)
+        try:
+            import librosa
+            audio, sr = librosa.load(file_path, sr=target_sr, mono=True)
+            # Audio is already in float32 format normalized to [-1, 1]
+            return audio.astype(np.float32)
+        except ImportError:
+            # Fallback to soundfile
+            import soundfile as sf
+            audio, sr = sf.read(file_path)
+            
+            # Convert stereo to mono if needed
+            if len(audio.shape) > 1:
+                audio = np.mean(audio, axis=1)
+            
+            # Resample if needed
+            if sr != target_sr:
+                import scipy.signal
+                num_samples = int(len(audio) * target_sr / sr)
+                audio = scipy.signal.resample(audio, num_samples)
+            
+            # Ensure audio is in the correct format (float32, normalized to [-1, 1])
+            if audio.dtype != np.float32:
+                if audio.dtype == np.int16:
+                    audio = audio.astype(np.float32) / 32768.0
+                elif audio.dtype == np.int32:
+                    audio = audio.astype(np.float32) / 2147483648.0
+                else:
+                    audio = audio.astype(np.float32)
+            
+            # Normalize to [-1, 1] if needed
+            if np.max(np.abs(audio)) > 1.0:
+                audio = audio / np.max(np.abs(audio))
+            
+            return audio.astype(np.float32)
+    except Exception as e:
+        print(f"Error loading {file_path}: {e}")
+        return None
+
+
+def extract_embedding_espnet(audio_path, speech2spk_embed):
+    """Extract embedding using ESPnet."""
+    try:
+        audio = load_audio(audio_path)
+        if audio is None:
+            return None
+        
+        # ESPnet expects numpy array (raw waveform, 16kHz, float32)
+        # The Speech2Embedding object handles the processing
+        embedding = speech2spk_embed(audio)
+        
+        # Convert to numpy array if needed (ESPnet may return numpy or torch tensor)
+        if hasattr(embedding, 'cpu'):  # torch.Tensor
+            embedding = embedding.cpu().numpy()
+        elif not isinstance(embedding, np.ndarray):
+            embedding = np.array(embedding)
+        
+        return embedding.flatten()
+    except Exception as e:
+        print(f"Error extracting embedding with ESPnet from {audio_path}: {e}")
+        import traceback
+        traceback.print_exc()
+        return None
+
+
+def main():
+    """Main function to extract embeddings for all audio files."""
+    
+    print("=" * 80)
+    print("RawNet3 Embedding Extraction")
+    print("=" * 80)
+    
+    # Initialize model
+    if not ESPNET_AVAILABLE:
+        print("\nERROR: ESPnet-SPK is not available.")
+        print("Please install ESPnet-SPK first:")
+        print("  pip install espnet")
+        print("\nThis script uses ESPnet-SPK's pre-trained RawNet3 model for easier access.")
+        sys.exit(1)
+    
+    print("\nInitializing ESPnet RawNet3 model...")
+    try:
+        speech2spk_embed = Speech2Embedding.from_pretrained(
+            model_tag="espnet/voxcelebs12_rawnet3"
+        )
+        print("ESPnet RawNet3 model loaded successfully")
+    except Exception as e:
+        print(f"\nERROR: Failed to load ESPnet RawNet3 model: {e}")
+        print("\nThis may be due to:")
+        print("  1. Network issues (model needs to be downloaded from HuggingFace)")
+        print("  2. Missing dependencies")
+        print("\nPlease ensure you have a stable internet connection and try again.")
+        import traceback
+        traceback.print_exc()
+        sys.exit(1)
+    
+    # Collect all audio files
+    print("\nCollecting audio files...")
+    
+    from extraction_utils import collect_audio_files, DEFAULT_OUTPUT_DIR
+    exp_2_files, output_files = collect_audio_files()
+    print(f"Found {len(exp_2_files)} reference files (data/audio/reference/*R.wav)")
+    print(f"Found {len(output_files)} comparison files (data/audio/comparison/*.wav)")
+    
+    total_files = len(exp_2_files) + len(output_files)
+    print(f"Total files to process: {total_files}")
+    
+    if total_files == 0:
+        print("ERROR: No audio files found!")
+        sys.exit(1)
+    
+    # Extract embeddings
+    print("\nExtracting embeddings...")
+    embeddings_dict = {}
+    failed_files = []
+    
+    # Process exp_2 files
+    print("\nProcessing exp_2 files...")
+    for audio_path in tqdm(exp_2_files, desc="exp_2"):
+        # Get filename without extension (e.g., "F01R")
+        filename_key = audio_path.stem
+        
+        try:
+            embedding = extract_embedding_espnet(audio_path, speech2spk_embed)
+            
+            if embedding is not None:
+                embeddings_dict[filename_key] = embedding
+            else:
+                failed_files.append(str(audio_path))
+        except Exception as e:
+            print(f"\nError processing {audio_path}: {e}")
+            failed_files.append(str(audio_path))
+    
+    # Process output files
+    print("\nProcessing output files...")
+    for audio_path in tqdm(output_files, desc="output"):
+        # Get filename without extension (e.g., "6_F01R_F02R_001")
+        filename_key = audio_path.stem
+        
+        try:
+            embedding = extract_embedding_espnet(audio_path, speech2spk_embed)
+            
+            if embedding is not None:
+                embeddings_dict[filename_key] = embedding
+            else:
+                failed_files.append(str(audio_path))
+        except Exception as e:
+            print(f"\nError processing {audio_path}: {e}")
+            failed_files.append(str(audio_path))
+    
+    # Save embeddings
+    print("\nSaving embeddings...")
+    
+    DEFAULT_OUTPUT_DIR.mkdir(parents=True, exist_ok=True)
+    npz_path = DEFAULT_OUTPUT_DIR / "rawnet3.npz"
+    np.savez_compressed(npz_path, **embeddings_dict)
+    print(f"Saved embeddings to {npz_path}")
+    
+    # Print summary
+    print("\n" + "=" * 80)
+    print("Summary")
+    print("=" * 80)
+    print(f"Total files processed: {total_files}")
+    print(f"Successfully extracted: {len(embeddings_dict)}")
+    print(f"Failed: {len(failed_files)}")
+    
+    if failed_files:
+        print(f"\nFailed files (first 10):")
+        for f in failed_files[:10]:
+            print(f"  {f}")
+        if len(failed_files) > 10:
+            print(f"  ... and {len(failed_files) - 10} more")
+    
+    print(f"\nEmbedding dimension: {list(embeddings_dict.values())[0].shape if embeddings_dict else 'N/A'}")
+    print(f"\nSaved files:")
+    print(f"  - {pickle_path} (Python pickle format)")
+    print(f"  - {npz_path} (NumPy compressed format)")
+    print("\nTo load embeddings later:")
+    print(f"  import pickle")
+    print(f"  with open('{pickle_path}', 'rb') as f:")
+    print(f"      embeddings = pickle.load(f)")
+    print("\nOr:")
+    print(f"  import numpy as np")
+    print(f"  data = np.load('{npz_path}', allow_pickle=True)")
+    print(f"  embeddings = {{k: data[k] for k in data.files}}")
+
+
+if __name__ == "__main__":
+    main()

code/extract_resemblyzer.py

@@ -0,0 +1,33 @@
+#!/usr/bin/env python3
+"""Extract Resemblyzer (GE2E) embeddings.
+
+Model: GE2E pre-trained speaker encoder (supervised, 256-dim)
+Install: pip install resemblyzer
+"""
+
+import argparse
+import numpy as np
+from extraction_utils import extract_all
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--base-dir", default=None)
+    parser.add_argument("--output-dir", default=None)
+    args = parser.parse_args()
+
+    from resemblyzer import VoiceEncoder, preprocess_wav
+
+    print("Loading Resemblyzer GE2E encoder...")
+    encoder = VoiceEncoder()
+
+    def model_fn(audio_path):
+        wav = preprocess_wav(str(audio_path))
+        embedding = encoder.embed_utterance(wav)  # (256,)
+        return embedding
+
+    extract_all(model_fn, "resemblyzer", args.base_dir, args.output_dir)
+
+
+if __name__ == "__main__":
+    main()

code/extract_ssl_layers.py

@@ -0,0 +1,131 @@
+#!/usr/bin/env python3
+"""Extract per-layer embeddings from SSL speech models.
+
+For wav2vec2, HuBERT, WavLM, XLS-R, and Whisper, the "last hidden state" used
+in the main extraction is known to underperform for speaker tasks relative to
+intermediate layers (see SUPERB benchmark). This script saves the mean-pooled
+embedding from EVERY transformer layer, enabling layer-wise analysis.
+
+Output format: for each audio file, a numpy array of shape (n_layers, dim).
+
+Usage:
+    python3 extract_ssl_layers.py --model wav2vec2 --device cpu
+    python3 extract_ssl_layers.py --model hubert --device cpu
+    python3 extract_ssl_layers.py --model wavlm --device cpu
+    python3 extract_ssl_layers.py --model xlsr --device cpu
+    python3 extract_ssl_layers.py --model whisper --device cpu
+"""
+
+import argparse
+import numpy as np
+import torch
+from extraction_utils import load_audio, extract_all
+
+
+HF_CONFIGS = {
+    'wav2vec2': {
+        'model_id': 'facebook/wav2vec2-base',
+        'model_cls': 'Wav2Vec2Model',
+        'fe_cls': 'Wav2Vec2FeatureExtractor',
+    },
+    'hubert': {
+        'model_id': 'facebook/hubert-base-ls960',
+        'model_cls': 'HubertModel',
+        'fe_cls': 'Wav2Vec2FeatureExtractor',
+    },
+    'wavlm': {
+        'model_id': 'microsoft/wavlm-base-plus',
+        'model_cls': 'WavLMModel',
+        'fe_cls': 'Wav2Vec2FeatureExtractor',
+    },
+    'xlsr': {
+        'model_id': 'facebook/wav2vec2-xls-r-300m',
+        'model_cls': 'Wav2Vec2Model',
+        'fe_cls': 'Wav2Vec2FeatureExtractor',
+    },
+}
+
+
+def build_hf_model_fn(model_name, device):
+    cfg = HF_CONFIGS[model_name]
+    import transformers
+    ModelCls = getattr(transformers, cfg['model_cls'])
+    FeCls = getattr(transformers, cfg['fe_cls'])
+    print(f"Loading {model_name} ({cfg['model_id']}) on {device}...")
+    feature_extractor = FeCls.from_pretrained(cfg['model_id'])
+    model = ModelCls.from_pretrained(cfg['model_id']).to(device)
+    model.eval()
+
+    def model_fn(audio_path):
+        audio = load_audio(audio_path, target_sr=16000)
+        inputs = feature_extractor(audio, sampling_rate=16000, return_tensors="pt")
+        inputs = {k: v.to(device) for k, v in inputs.items()}
+        with torch.no_grad():
+            outputs = model(**inputs, output_hidden_states=True)
+        # outputs.hidden_states: tuple of (n_layers+1) tensors, each (1, T, D)
+        # Index 0 is the CNN feature projection output; indices 1..n are transformer layers
+        all_layers = torch.stack(outputs.hidden_states, dim=0)  # (L+1, 1, T, D)
+        # Mean pool over time for each layer
+        pooled = all_layers.mean(dim=2).squeeze(1)  # (L+1, D)
+        return pooled.cpu().numpy()
+
+    return model_fn
+
+
+def build_whisper_model_fn(device, size='base'):
+    import whisper
+    print(f"Loading Whisper {size} encoder on {device}...")
+    model = whisper.load_model(size, device=device)
+
+    def model_fn(audio_path):
+        audio = whisper.load_audio(str(audio_path))
+        audio = whisper.pad_or_trim(audio)
+        mel = whisper.log_mel_spectrogram(audio).to(device)
+        # Whisper's encoder returns only the final output by default.
+        # To get per-layer outputs, register a forward hook.
+        layer_outputs = []
+
+        def hook(module, input_, output):
+            # output is typically a tensor (B, T, D) or tuple
+            out = output[0] if isinstance(output, tuple) else output
+            layer_outputs.append(out.mean(dim=1).squeeze().cpu().numpy())  # (D,)
+
+        handles = []
+        for block in model.encoder.blocks:
+            handles.append(block.register_forward_hook(hook))
+
+        with torch.no_grad():
+            _ = model.encoder(mel.unsqueeze(0))
+
+        for h in handles:
+            h.remove()
+
+        # Also include final layer norm output
+        with torch.no_grad():
+            final = model.encoder(mel.unsqueeze(0)).mean(dim=1).squeeze().cpu().numpy()
+        layer_outputs.append(final)
+
+        return np.stack(layer_outputs, axis=0)  # (L, D)
+
+    return model_fn
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model", required=True,
+                        choices=['wav2vec2', 'hubert', 'wavlm', 'xlsr', 'whisper'])
+    parser.add_argument("--device", default="cpu")
+    parser.add_argument("--base-dir", default=None)
+    parser.add_argument("--output-dir", default=None)
+    args = parser.parse_args()
+
+    if args.model == 'whisper':
+        model_fn = build_whisper_model_fn(args.device)
+    else:
+        model_fn = build_hf_model_fn(args.model, args.device)
+
+    extract_all(model_fn, f"{args.model}_layers", args.base_dir, args.output_dir)
+
+
+if __name__ == "__main__":
+    main()

code/extract_titanet.py

@@ -0,0 +1,39 @@
+#!/usr/bin/env python3
+"""Extract TitaNet embeddings using NVIDIA NeMo.
+
+Model: nvidia/speakerverification_en_titanet_large (supervised, 192-dim)
+Install: pip install nemo_toolkit[asr]
+"""
+
+import argparse
+import torch
+import numpy as np
+from extraction_utils import load_audio, extract_all
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--device", default="cuda" if torch.cuda.is_available() else "cpu")
+    parser.add_argument("--base-dir", default=None)
+    parser.add_argument("--output-dir", default=None)
+    args = parser.parse_args()
+
+    import nemo.collections.asr as nemo_asr
+
+    print(f"Loading TitaNet on {args.device}...")
+    model = nemo_asr.models.EncDecSpeakerLabelModel.from_pretrained(
+        "nvidia/speakerverification_en_titanet_large"
+    )
+    model = model.to(args.device)
+    model.eval()
+
+    def model_fn(audio_path):
+        # NeMo's get_embedding works directly with file paths
+        emb = model.get_embedding(str(audio_path))
+        return emb.squeeze().cpu().numpy()
+
+    extract_all(model_fn, "titanet", args.base_dir, args.output_dir)
+
+
+if __name__ == "__main__":
+    main()

code/extract_wav2vec2.py

@@ -0,0 +1,45 @@
+#!/usr/bin/env python3
+"""Extract wav2vec 2.0 embeddings using HuggingFace Transformers.
+
+Model: facebook/wav2vec2-base (self-supervised ONLY, 768-dim)
+  - Pre-trained with contrastive loss on LibriSpeech 960h (Baevski et al., NeurIPS 2020)
+  - NOT the ASR-fine-tuned version (wav2vec2-base-960h has CTC fine-tuning)
+Frame-level output is mean-pooled to get utterance-level embeddings.
+Install: pip install transformers torchaudio
+"""
+
+import argparse
+import torch
+import numpy as np
+from extraction_utils import load_audio, extract_all
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--device", default="cuda" if torch.cuda.is_available() else "cpu")
+    parser.add_argument("--base-dir", default=None)
+    parser.add_argument("--output-dir", default=None)
+    args = parser.parse_args()
+
+    from transformers import Wav2Vec2Model, Wav2Vec2FeatureExtractor
+
+    print(f"Loading wav2vec2-base (self-supervised, no ASR fine-tuning) on {args.device}...")
+    feature_extractor = Wav2Vec2FeatureExtractor.from_pretrained("facebook/wav2vec2-base")
+    model = Wav2Vec2Model.from_pretrained("facebook/wav2vec2-base").to(args.device)
+    model.eval()
+
+    def model_fn(audio_path):
+        audio = load_audio(audio_path, target_sr=16000)
+        inputs = feature_extractor(audio, sampling_rate=16000, return_tensors="pt")
+        inputs = {k: v.to(args.device) for k, v in inputs.items()}
+        with torch.no_grad():
+            outputs = model(**inputs)
+        hidden = outputs.last_hidden_state  # (1, T, 768)
+        embedding = hidden.mean(dim=1).squeeze().cpu().numpy()  # (768,)
+        return embedding
+
+    extract_all(model_fn, "wav2vec2", args.base_dir, args.output_dir)
+
+
+if __name__ == "__main__":
+    main()

code/extract_wavlm.py

@@ -0,0 +1,43 @@
+#!/usr/bin/env python3
+"""Extract WavLM embeddings using HuggingFace Transformers.
+
+Model: microsoft/wavlm-base-plus (self-supervised, 768-dim)
+Frame-level output is mean-pooled to get utterance-level embeddings.
+Install: pip install transformers
+"""
+
+import argparse
+import torch
+import numpy as np
+from extraction_utils import load_audio, extract_all
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--device", default="cuda" if torch.cuda.is_available() else "cpu")
+    parser.add_argument("--base-dir", default=None)
+    parser.add_argument("--output-dir", default=None)
+    args = parser.parse_args()
+
+    from transformers import WavLMModel, Wav2Vec2FeatureExtractor
+
+    print(f"Loading WavLM-base-plus on {args.device}...")
+    feature_extractor = Wav2Vec2FeatureExtractor.from_pretrained("microsoft/wavlm-base-plus")
+    model = WavLMModel.from_pretrained("microsoft/wavlm-base-plus").to(args.device)
+    model.eval()
+
+    def model_fn(audio_path):
+        audio = load_audio(audio_path, target_sr=16000)
+        inputs = feature_extractor(audio, sampling_rate=16000, return_tensors="pt")
+        inputs = {k: v.to(args.device) for k, v in inputs.items()}
+        with torch.no_grad():
+            outputs = model(**inputs)
+        hidden = outputs.last_hidden_state  # (1, T, 768)
+        embedding = hidden.mean(dim=1).squeeze().cpu().numpy()
+        return embedding
+
+    extract_all(model_fn, "wavlm", args.base_dir, args.output_dir)
+
+
+if __name__ == "__main__":
+    main()

code/extract_whisper.py

@@ -0,0 +1,41 @@
+#!/usr/bin/env python3
+"""Extract Whisper encoder embeddings.
+
+Model: openai/whisper-base (encoder only, 512-dim after mean pooling)
+Frame-level encoder output is mean-pooled to get utterance-level embeddings.
+Install: pip install openai-whisper
+"""
+
+import argparse
+import torch
+import numpy as np
+from extraction_utils import extract_all
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--device", default="cuda" if torch.cuda.is_available() else "cpu")
+    parser.add_argument("--model-size", default="base", choices=["tiny", "base", "small", "medium"])
+    parser.add_argument("--base-dir", default=None)
+    parser.add_argument("--output-dir", default=None)
+    args = parser.parse_args()
+
+    import whisper
+
+    print(f"Loading Whisper {args.model_size} on {args.device}...")
+    model = whisper.load_model(args.model_size, device=args.device)
+
+    def model_fn(audio_path):
+        audio = whisper.load_audio(str(audio_path))
+        audio = whisper.pad_or_trim(audio)
+        mel = whisper.log_mel_spectrogram(audio).to(args.device)
+        with torch.no_grad():
+            enc_output = model.encoder(mel.unsqueeze(0))  # (1, T, D)
+        embedding = enc_output.mean(dim=1).squeeze().cpu().numpy()
+        return embedding
+
+    extract_all(model_fn, "whisper", args.base_dir, args.output_dir)
+
+
+if __name__ == "__main__":
+    main()

code/extract_xlsr.py

@@ -0,0 +1,43 @@
+#!/usr/bin/env python3
+"""Extract XLS-R (wav2vec2 multilingual) embeddings.
+
+Model: facebook/wav2vec2-xls-r-300m (self-supervised multilingual, 1024-dim)
+Frame-level output is mean-pooled to get utterance-level embeddings.
+Install: pip install transformers
+"""
+
+import argparse
+import torch
+import numpy as np
+from extraction_utils import load_audio, extract_all
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--device", default="cuda" if torch.cuda.is_available() else "cpu")
+    parser.add_argument("--base-dir", default=None)
+    parser.add_argument("--output-dir", default=None)
+    args = parser.parse_args()
+
+    from transformers import Wav2Vec2Model, Wav2Vec2FeatureExtractor
+
+    print(f"Loading XLS-R 300M on {args.device}...")
+    feature_extractor = Wav2Vec2FeatureExtractor.from_pretrained("facebook/wav2vec2-xls-r-300m")
+    model = Wav2Vec2Model.from_pretrained("facebook/wav2vec2-xls-r-300m").to(args.device)
+    model.eval()
+
+    def model_fn(audio_path):
+        audio = load_audio(audio_path, target_sr=16000)
+        inputs = feature_extractor(audio, sampling_rate=16000, return_tensors="pt")
+        inputs = {k: v.to(args.device) for k, v in inputs.items()}
+        with torch.no_grad():
+            outputs = model(**inputs)
+        hidden = outputs.last_hidden_state  # (1, T, 1024)
+        embedding = hidden.mean(dim=1).squeeze().cpu().numpy()
+        return embedding
+
+    extract_all(model_fn, "xlsr", args.base_dir, args.output_dir)
+
+
+if __name__ == "__main__":
+    main()

code/extract_xvector.py

@@ -0,0 +1,42 @@
+#!/usr/bin/env python3
+"""Extract x-vector embeddings using SpeechBrain.
+
+Model: speechbrain/spkrec-xvect-voxceleb (supervised, softmax, 512-dim)
+  - The classic TDNN x-vector architecture (Snyder et al., ICASSP 2018)
+  - Trained on VoxCeleb1+2 with softmax loss
+  - Statistics pooling (mean + std) for utterance-level embeddings
+Install: pip install speechbrain
+"""
+
+import argparse
+import torch
+import numpy as np
+from extraction_utils import load_audio, extract_all
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--device", default="cuda" if torch.cuda.is_available() else "cpu")
+    parser.add_argument("--base-dir", default=None)
+    parser.add_argument("--output-dir", default=None)
+    args = parser.parse_args()
+
+    from speechbrain.inference.speaker import EncoderClassifier
+
+    print(f"Loading x-vector on {args.device}...")
+    classifier = EncoderClassifier.from_hparams(
+        source="speechbrain/spkrec-xvect-voxceleb",
+        run_opts={"device": args.device},
+    )
+
+    def model_fn(audio_path):
+        audio = load_audio(audio_path, target_sr=16000)
+        signal = torch.tensor(audio).unsqueeze(0).to(args.device)
+        embedding = classifier.encode_batch(signal)
+        return embedding.squeeze().cpu().numpy()
+
+    extract_all(model_fn, "xvector", args.base_dir, args.output_dir)
+
+
+if __name__ == "__main__":
+    main()

code/extraction_utils.py

@@ -0,0 +1,167 @@
+"""
+Shared utilities for embedding extraction across all models.
+
+Provides common audio loading, file collection, and save logic so that
+each per-model extraction script only needs to define model initialization
+and a single `model_fn(audio_path) -> np.ndarray` function.
+"""
+
+import sys
+import numpy as np
+from pathlib import Path
+
+try:
+    from tqdm import tqdm
+except ImportError:
+    def tqdm(iterable, desc=None, total=None):
+        if desc:
+            print(f"\nProcessing {desc}...")
+        return iterable
+
+import os
+
+# Resolve the release root.
+# Priority: VIPBENCH_ROOT env var > parent of this script's directory.
+# Layout assumed:
+#   <VIPBENCH_ROOT>/
+#     code/extraction_utils.py    <- this file
+#     data/audio/reference/*.wav
+#     data/audio/comparison/*.wav
+#     data/embeddings/<model>.npz <- output
+def _resolve_root():
+    env = os.environ.get("VIPBENCH_ROOT")
+    if env:
+        return Path(env).resolve()
+    return Path(__file__).resolve().parent.parent
+
+DEFAULT_BASE_DIR = _resolve_root()
+DEFAULT_OUTPUT_DIR = DEFAULT_BASE_DIR / "data" / "embeddings"
+
+
+def load_audio(file_path, target_sr=16000):
+    """Load audio file as float32 mono at target_sr.
+
+    Replicates the pattern from extract_rawnet3_embeddings.py.
+    """
+    try:
+        import librosa
+        audio, sr = librosa.load(file_path, sr=target_sr, mono=True)
+        return audio.astype(np.float32)
+    except ImportError:
+        pass
+
+    # Fallback to soundfile
+    import soundfile as sf
+    audio, sr = sf.read(file_path)
+    if len(audio.shape) > 1:
+        audio = np.mean(audio, axis=1)
+    if sr != target_sr:
+        import scipy.signal
+        num_samples = int(len(audio) * target_sr / sr)
+        audio = scipy.signal.resample(audio, num_samples)
+    if audio.dtype != np.float32:
+        if audio.dtype == np.int16:
+            audio = audio.astype(np.float32) / 32768.0
+        elif audio.dtype == np.int32:
+            audio = audio.astype(np.float32) / 2147483648.0
+        else:
+            audio = audio.astype(np.float32)
+    if np.max(np.abs(audio)) > 1.0:
+        audio = audio / np.max(np.abs(audio))
+    return audio.astype(np.float32)
+
+
+def collect_audio_files(base_dir=None):
+    """Collect reference and stimulus audio files.
+
+    Returns
+    -------
+    ref_files : list[Path]
+        100 reference clips from exp_2/*R.wav
+    stim_files : list[Path]
+        9,800 comparison clips from output/*.wav
+    """
+    base = Path(base_dir) if base_dir else DEFAULT_BASE_DIR
+    ref_files = sorted((base / "data" / "audio" / "reference").glob("*R.wav"))
+    stim_files = sorted((base / "data" / "audio" / "comparison").glob("*.wav"))
+    return ref_files, stim_files
+
+
+def save_embeddings(embeddings_dict, model_name, output_dir=None):
+    """Save embeddings as compressed .npz with same format as rawnet3_embeddings.npz."""
+    out = Path(output_dir) if output_dir else DEFAULT_OUTPUT_DIR
+    out.mkdir(parents=True, exist_ok=True)
+    path = out / f"{model_name}.npz"
+    np.savez_compressed(path, **embeddings_dict)
+    print(f"Saved {len(embeddings_dict)} embeddings to {path}")
+    return path
+
+
+def extract_all(model_fn, model_name, base_dir=None, output_dir=None):
+    """Run extraction for all audio files using the provided model function.
+
+    Parameters
+    ----------
+    model_fn : callable
+        Takes a file path (str or Path) and returns a 1-D numpy array (the embedding).
+    model_name : str
+        Name used for the output file ({model_name}_embeddings.npz).
+    base_dir : str or Path, optional
+        Project root. Defaults to DEFAULT_BASE_DIR.
+    output_dir : str or Path, optional
+        Where to save the .npz. Defaults to DEFAULT_OUTPUT_DIR.
+    """
+    ref_files, stim_files = collect_audio_files(base_dir)
+    total = len(ref_files) + len(stim_files)
+    print(f"\n{'=' * 60}")
+    print(f"{model_name} Embedding Extraction")
+    print(f"{'=' * 60}")
+    print(f"Reference files: {len(ref_files)}")
+    print(f"Stimulus files:  {len(stim_files)}")
+    print(f"Total:           {total}")
+
+    embeddings_dict = {}
+    failed = []
+
+    print(f"\nProcessing reference files...")
+    for path in tqdm(ref_files, desc="references"):
+        key = path.stem
+        try:
+            emb = model_fn(path)
+            if emb is not None:
+                embeddings_dict[key] = emb
+            else:
+                failed.append(str(path))
+        except Exception as e:
+            print(f"\n  Error on {path.name}: {e}")
+            failed.append(str(path))
+
+    print(f"\nProcessing stimulus files...")
+    for path in tqdm(stim_files, desc="stimuli"):
+        key = path.stem
+        try:
+            emb = model_fn(path)
+            if emb is not None:
+                embeddings_dict[key] = emb
+            else:
+                failed.append(str(path))
+        except Exception as e:
+            print(f"\n  Error on {path.name}: {e}")
+            failed.append(str(path))
+
+    # Summary
+    print(f"\n{'=' * 60}")
+    print(f"Summary")
+    print(f"{'=' * 60}")
+    print(f"Extracted: {len(embeddings_dict)} / {total}")
+    print(f"Failed:    {len(failed)}")
+    if embeddings_dict:
+        sample = next(iter(embeddings_dict.values()))
+        print(f"Dimension: {sample.shape}")
+    if failed:
+        print(f"\nFailed files (first 10):")
+        for f in failed[:10]:
+            print(f"  {f}")
+
+    save_embeddings(embeddings_dict, model_name, output_dir)
+    return embeddings_dict

code/reproduce.sh

@@ -0,0 +1,33 @@
+#!/usr/bin/env bash
+# VIPBench end-to-end reproduction.
+#
+# Default: skip extraction (use the embeddings shipped in data/embeddings/) and
+# execute the analysis notebook end-to-end. ~10 minutes on a single laptop.
+#
+# To re-extract embeddings from raw audio (~24 CPU-hours plus ~1 GPU-hour for
+# Whisper), pass `--extract` as the first argument.
+#
+# Usage:
+#   bash reproduce.sh              # analysis only, uses shipped embeddings
+#   bash reproduce.sh --extract    # re-extract everything from audio first
+#   VIPBENCH_ROOT=/path/to/release bash reproduce.sh   # explicit root override
+set -euo pipefail
+
+cd "$(dirname "$0")"
+export VIPBENCH_ROOT="${VIPBENCH_ROOT:-$(pwd)/..}"
+echo "VIPBENCH_ROOT=${VIPBENCH_ROOT}"
+
+if [[ "${1:-}" == "--extract" ]]; then
+  echo "[1/2] Re-extracting embeddings (this will take many hours)..."
+  bash run_all_extractions.sh
+else
+  echo "[1/2] Skipping extraction. Using embeddings in data/embeddings/."
+  echo "      Pass --extract to re-extract from audio."
+fi
+
+echo "[2/2] Executing analysis notebook..."
+jupyter nbconvert --to notebook --execute benchmark_analysis.ipynb \
+  --output benchmark_analysis_executed.ipynb \
+  --ExecutePreprocessor.timeout=1800
+
+echo "Done. See code/benchmark_analysis_executed.ipynb for the rendered output."

code/run_all_extractions.sh

@@ -0,0 +1,55 @@
+#!/bin/bash
+# Run all embedding extraction scripts sequentially.
+# Each script saves <model>.npz to <VIPBENCH_ROOT>/data/embeddings/.
+# Scripts that fail (e.g., missing packages) are skipped gracefully.
+
+set -o pipefail
+cd "$(dirname "$0")"
+export VIPBENCH_ROOT="${VIPBENCH_ROOT:-$(pwd)/..}"
+echo "VIPBENCH_ROOT=${VIPBENCH_ROOT}"
+mkdir -p "${VIPBENCH_ROOT}/data/embeddings"
+
+echo "============================================================"
+echo "Embedding Extraction Pipeline"
+echo "============================================================"
+echo ""
+
+# Run each extractor, ordered from lightest to heaviest install
+SCRIPTS=(
+    extract_rawnet3_embeddings.py
+    extract_whisper.py
+    extract_resemblyzer.py
+    extract_ecapa_tdnn.py
+    extract_xvector.py
+    extract_wav2vec2.py
+    extract_hubert.py
+    extract_wavlm.py
+    extract_xlsr.py
+    extract_titanet.py
+)
+
+SUCCEEDED=0
+FAILED=0
+
+for script in "${SCRIPTS[@]}"; do
+    echo "============================================================"
+    echo "Running: $script"
+    echo "============================================================"
+    if python3 "$script" 2>&1; then
+        echo "  -> SUCCESS"
+        SUCCEEDED=$((SUCCEEDED + 1))
+    else
+        echo "  -> FAILED (continuing...)"
+        FAILED=$((FAILED + 1))
+    fi
+    echo ""
+done
+
+echo "============================================================"
+echo "Pipeline Complete"
+echo "  Succeeded: $SUCCEEDED"
+echo "  Failed:    $FAILED"
+echo "============================================================"
+echo ""
+echo "Available embeddings:"
+ls -lh "${VIPBENCH_ROOT}/data/embeddings/"*.npz 2>/dev/null || echo "  (none found)"

croissant.json

@@ -0,0 +1,445 @@
+{
+  "@context": {
+    "@language": "en",
+    "@vocab": "https://schema.org/",
+    "citeAs": "cr:citeAs",
+    "column": "cr:column",
+    "conformsTo": "dct:conformsTo",
+    "cr": "http://mlcommons.org/croissant/",
+    "rai": "http://mlcommons.org/croissant/RAI/",
+    "data": {
+      "@id": "cr:data",
+      "@type": "@json"
+    },
+    "dataType": {
+      "@id": "cr:dataType",
+      "@type": "@vocab"
+    },
+    "dct": "http://purl.org/dc/terms/",
+    "examples": {
+      "@id": "cr:examples",
+      "@type": "@json"
+    },
+    "extract": "cr:extract",
+    "field": "cr:field",
+    "fileProperty": "cr:fileProperty",
+    "fileObject": "cr:fileObject",
+    "fileSet": "cr:fileSet",
+    "format": "cr:format",
+    "includes": "cr:includes",
+    "isLiveDataset": "cr:isLiveDataset",
+    "jsonPath": "cr:jsonPath",
+    "key": "cr:key",
+    "md5": "cr:md5",
+    "parentField": "cr:parentField",
+    "path": "cr:path",
+    "recordSet": "cr:recordSet",
+    "references": "cr:references",
+    "regex": "cr:regex",
+    "repeated": "cr:repeated",
+    "replace": "cr:replace",
+    "sc": "https://schema.org/",
+    "separator": "cr:separator",
+    "source": "cr:source",
+    "subField": "cr:subField",
+    "transform": "cr:transform"
+  },
+  "@type": "sc:Dataset",
+  "name": "VIPBench",
+  "description": "A human-aligned benchmark for voice identity perception. 124,876 same/different identity judgments from 1,290 English-speaking listeners on 9,800 voice pairs across 100 demographically-stratified speakers, spanning real recordings, AI voice clones, and continuously morphed voices. Includes pre-extracted embeddings for 10 speaker and speech-representation models.",
+  "conformsTo": "http://mlcommons.org/croissant/1.0",
+  "citeAs": "Anonymous Authors. VIPBench: A Human-Aligned Benchmark for Voice Identity Perception in the Age of Voice Cloning. Advances in Neural Information Processing Systems Datasets and Benchmarks Track, 2026. Anonymized for double-blind review.",
+  "license": "https://creativecommons.org/licenses/by-nc/4.0/",
+  "url": "https://huggingface.co/datasets/anonymous/vipbench",
+  "version": "1.0.0",
+  "datePublished": "2026-05-06",
+  "keywords": [
+    "speaker embeddings",
+    "voice identity perception",
+    "human-aligned benchmark",
+    "voice cloning",
+    "perceptual evaluation",
+    "speaker verification",
+    "speech representations"
+  ],
+  "creator": {
+    "@type": "Organization",
+    "name": "Anonymous (NeurIPS 2026 double-blind review)"
+  },
+  "isLiveDataset": false,
+  "distribution": [
+    {
+      "@type": "cr:FileObject",
+      "@id": "speakers-csv",
+      "name": "speakers.csv",
+      "description": "Per-speaker metadata for the 100 celebrity speakers.",
+      "contentUrl": "data/speakers.csv",
+      "encodingFormat": "text/csv",
+      "sha256": "TO_BE_COMPUTED_BY_PUBLISHER"
+    },
+    {
+      "@type": "cr:FileObject",
+      "@id": "stimuli-csv",
+      "name": "stimuli.csv",
+      "description": "Per-pair aggregates for the 9,800 voice pairs.",
+      "contentUrl": "data/stimuli.csv",
+      "encodingFormat": "text/csv",
+      "sha256": "TO_BE_COMPUTED_BY_PUBLISHER"
+    },
+    {
+      "@type": "cr:FileObject",
+      "@id": "responses-csv",
+      "name": "participant_responses.csv",
+      "description": "Per-judgment records (124,876 rows).",
+      "contentUrl": "data/participant_responses.csv",
+      "encodingFormat": "text/csv",
+      "sha256": "TO_BE_COMPUTED_BY_PUBLISHER"
+    },
+    {
+      "@type": "cr:FileObject",
+      "@id": "stimuli-interpol-csv",
+      "name": "stimuli_interpol.csv",
+      "description": "Type 6 morph trajectory metadata (8,100 rows).",
+      "contentUrl": "data/stimuli_interpol.csv",
+      "encodingFormat": "text/csv",
+      "sha256": "TO_BE_COMPUTED_BY_PUBLISHER"
+    },
+    {
+      "@type": "cr:FileSet",
+      "@id": "audio-reference",
+      "name": "audio/reference",
+      "description": "100 reference audio clips (16 kHz mono WAV), one per speaker.",
+      "containedIn": {
+        "@id": "vipbench-bundle"
+      },
+      "encodingFormat": "audio/wav",
+      "includes": "data/audio/reference/*.wav"
+    },
+    {
+      "@type": "cr:FileSet",
+      "@id": "audio-comparison",
+      "name": "audio/comparison",
+      "description": "9,800 comparison audio clips (16 kHz mono WAV).",
+      "containedIn": {
+        "@id": "vipbench-bundle"
+      },
+      "encodingFormat": "audio/wav",
+      "includes": "data/audio/comparison/*.wav"
+    },
+    {
+      "@type": "cr:FileSet",
+      "@id": "embeddings-main",
+      "name": "embeddings",
+      "description": "Pre-extracted utterance-level embeddings for 10 speaker and speech-representation models. Each .npz contains 9,900 keys (audio basenames) mapped to 1-D embedding vectors.",
+      "containedIn": {
+        "@id": "vipbench-bundle"
+      },
+      "encodingFormat": "application/x-npz",
+      "includes": "data/embeddings/*.npz"
+    },
+    {
+      "@type": "cr:FileSet",
+      "@id": "embeddings-layers",
+      "name": "embeddings/layers",
+      "description": "Per-transformer-layer mean-pooled embeddings for 5 SSL models (wav2vec 2.0, HuBERT, WavLM, XLS-R, Whisper). Each .npz contains 9,900 keys mapped to 2-D arrays of shape (num_layers, embedding_dim).",
+      "containedIn": {
+        "@id": "vipbench-bundle"
+      },
+      "encodingFormat": "application/x-npz",
+      "includes": "data/embeddings/layers/*.npz"
+    }
+  ],
+  "recordSet": [
+    {
+      "@type": "cr:RecordSet",
+      "@id": "speakers",
+      "name": "speakers",
+      "description": "100 celebrity speakers stratified across 5 sociophonetic groups x 2 genders x 2 age brackets.",
+      "field": [
+        {
+          "@type": "cr:Field",
+          "@id": "speakers/id",
+          "name": "id",
+          "description": "Speaker identifier (e.g., F01, M07).",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {"@id": "speakers-csv"},
+            "extract": {"column": "id"}
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "speakers/name",
+          "name": "name",
+          "description": "Speaker name (public figure).",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {"@id": "speakers-csv"},
+            "extract": {"column": "name"}
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "speakers/group",
+          "name": "group",
+          "description": "Sociophonetic group (1=Italian-American, 2=African-American, 3=Asian-American, 4=Latinx, 5=mainstream Standard American English).",
+          "dataType": "sc:Integer",
+          "source": {
+            "fileObject": {"@id": "speakers-csv"},
+            "extract": {"column": "group"}
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "speakers/gender",
+          "name": "gender",
+          "description": "Speaker gender (1=male, 2=female).",
+          "dataType": "sc:Integer",
+          "source": {
+            "fileObject": {"@id": "speakers-csv"},
+            "extract": {"column": "gender"}
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "speakers/age",
+          "name": "age",
+          "description": "Speaker age bracket (1=younger, 2=older).",
+          "dataType": "sc:Integer",
+          "source": {
+            "fileObject": {"@id": "speakers-csv"},
+            "extract": {"column": "age"}
+          }
+        }
+      ]
+    },
+    {
+      "@type": "cr:RecordSet",
+      "@id": "stimuli",
+      "name": "stimuli",
+      "description": "9,800 voice pairs with per-pair human-judgment aggregates and metadata.",
+      "field": [
+        {
+          "@type": "cr:Field",
+          "@id": "stimuli/id",
+          "name": "id",
+          "description": "Stimulus identifier (matches the audio basename for the comparison clip).",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {"@id": "stimuli-csv"},
+            "extract": {"column": "id"}
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "stimuli/stimuli_type",
+          "name": "stimuli_type",
+          "description": "Stimulus type (1-6). See docs/stimulus_types.md.",
+          "dataType": "sc:Integer",
+          "source": {
+            "fileObject": {"@id": "stimuli-csv"},
+            "extract": {"column": "stimuli_type"}
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "stimuli/reference",
+          "name": "reference",
+          "description": "Reference speaker ID (joins to speakers/id).",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {"@id": "stimuli-csv"},
+            "extract": {"column": "reference"}
+          },
+          "references": {"field": {"@id": "speakers/id"}}
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "stimuli/comparison",
+          "name": "comparison",
+          "description": "Comparison speaker ID for non-Type-6 pairs.",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {"@id": "stimuli-csv"},
+            "extract": {"column": "comparison"}
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "stimuli/voice_clone",
+          "name": "voice_clone",
+          "description": "Whether the comparison clip is an AI voice clone (1) or natural recording (0).",
+          "dataType": "sc:Integer",
+          "source": {
+            "fileObject": {"@id": "stimuli-csv"},
+            "extract": {"column": "voice_clone"}
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "stimuli/correct_answer",
+          "name": "correct_answer",
+          "description": "Metadata-label same/different (1=same speaker by metadata, 0=different).",
+          "dataType": "sc:Integer",
+          "source": {
+            "fileObject": {"@id": "stimuli-csv"},
+            "extract": {"column": "correct_answer"}
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "stimuli/scale",
+          "name": "scale",
+          "description": "For Type 6 morphs, the interpolation scale (0=reference voice, 100=other voice). 100 for non-morph pairs.",
+          "dataType": "sc:Integer",
+          "source": {
+            "fileObject": {"@id": "stimuli-csv"},
+            "extract": {"column": "scale"}
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "stimuli/num_response",
+          "name": "num_response",
+          "description": "Number of listener judgments collected for this pair.",
+          "dataType": "sc:Integer",
+          "source": {
+            "fileObject": {"@id": "stimuli-csv"},
+            "extract": {"column": "num_response"}
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "stimuli/same_vote",
+          "name": "same_vote",
+          "description": "Number of listeners who judged the pair as the same speaker.",
+          "dataType": "sc:Integer",
+          "source": {
+            "fileObject": {"@id": "stimuli-csv"},
+            "extract": {"column": "same_vote"}
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "stimuli/diff_vote",
+          "name": "diff_vote",
+          "description": "Number of listeners who judged the pair as different speakers.",
+          "dataType": "sc:Integer",
+          "source": {
+            "fileObject": {"@id": "stimuli-csv"},
+            "extract": {"column": "diff_vote"}
+          }
+        }
+      ]
+    },
+    {
+      "@type": "cr:RecordSet",
+      "@id": "responses",
+      "name": "participant_responses",
+      "description": "124,876 individual listener judgments.",
+      "field": [
+        {
+          "@type": "cr:Field",
+          "@id": "responses/user_id",
+          "name": "user_id",
+          "description": "Pseudonymized listener identifier.",
+          "dataType": "sc:Integer",
+          "source": {
+            "fileObject": {"@id": "responses-csv"},
+            "extract": {"column": "user_id"}
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "responses/stimuli_id",
+          "name": "stimuli_id",
+          "description": "Stimulus identifier (joins to stimuli/id).",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {"@id": "responses-csv"},
+            "extract": {"column": "stimuli_id"}
+          },
+          "references": {"field": {"@id": "stimuli/id"}}
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "responses/stimuli_type",
+          "name": "stimuli_type",
+          "description": "Stimulus type (1-6).",
+          "dataType": "sc:Integer",
+          "source": {
+            "fileObject": {"@id": "responses-csv"},
+            "extract": {"column": "stimuli_type"}
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "responses/answer",
+          "name": "answer",
+          "description": "Listener's binary same/different judgment (1=same, 0=different).",
+          "dataType": "sc:Integer",
+          "source": {
+            "fileObject": {"@id": "responses-csv"},
+            "extract": {"column": "answer"}
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "responses/correct",
+          "name": "correct",
+          "description": "Whether the listener's answer matches the metadata label.",
+          "dataType": "sc:Integer",
+          "source": {
+            "fileObject": {"@id": "responses-csv"},
+            "extract": {"column": "correct"}
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "responses/know_speaker",
+          "name": "know_speaker",
+          "description": "Listener-recognition flag (1=listener identified the reference speaker).",
+          "dataType": "sc:Integer",
+          "source": {
+            "fileObject": {"@id": "responses-csv"},
+            "extract": {"column": "know_speaker"}
+          }
+        }
+      ]
+    }
+  ],
+  "rai:dataCollection": "Listener judgments were collected via the Centaur AI online crowdsourcing platform under an Institutional Review Board (IRB) approved research protocol. Each pair received at least 8 judgments from English-speaking adult participants. Source audio was selected from publicly available recordings of 100 US celebrity speakers; AI voice clones were generated from these source clips using a state-of-the-art text-to-speech system; voice morphs were generated by interpolating a voice-conversion latent between two source speakers.",
+  "rai:dataCollectionType": [
+    "Crowdsourcing",
+    "Synthetic data generation"
+  ],
+  "rai:dataCollectionMissingValues": "The know_speaker field is missing for some early-trial responses (less than 1% of records). Listeners with fewer than the qualification threshold of attention-check passes are flagged but their responses are still released.",
+  "rai:dataCollectionTimeFrame": {
+    "@type": "DateTime",
+    "@value": "2025-01-01/2026-04-30"
+  },
+  "rai:dataAnnotationProtocol": "Each trial presents a single audio clip in which a reference recording is followed by 1 second of silence, a short beep, and a comparison recording. The listener answers (a) whether the two clips came from the same speaker (binary same/different) and (b) optionally identifies which of four within-group celebrities (or 'I don't know') they recognize in the reference clip. Stimulus presentation order is randomized within participant. Participants provide informed consent prior to participating; no deception is involved.",
+  "rai:dataAnnotationPlatform": "Centaur AI crowdsourcing platform (https://centaur.ai)",
+  "rai:dataAnnotationAnalysis": "Per-pair P(same) is computed as the fraction of listeners who judged the pair as the same speaker. Inter-rater agreement is reported as Spearman-Brown corrected split-half reliability (rho_SB = 0.705 over 100 random splits).",
+  "rai:annotationsPerItem": "median 10, range 8 to 92",
+  "rai:dataPreprocessingProtocol": "Audio was resampled to 16 kHz mono. Listener responses were validated against attention-check probes embedded in the study. The reorganized_stimuli.csv aggregates count same/different votes per pair.",
+  "rai:dataReleaseMaintenancePlan": "The dataset is versioned (v1.0 at NeurIPS 2026 submission). Updates are tracked in CHANGELOG.md and tagged on the public repository. The authors will respond to errata via the public repository issue tracker (link to be added at camera-ready).",
+  "rai:dataLifecycleStage": "Released",
+  "rai:personalSensitiveInformation": "The dataset names public-figure speakers because the celebrity-stratified design is integral to the benchmark and source recordings are publicly available. Listener identifiers are pseudonymized integers tied to no external account. Listener demographic fields are limited to age band, gender, and a binary first-language flag.",
+  "rai:dataSocialImpact": "Voice-cloning systems that better align with human perception could inform adversarial use such as more convincing fraudulent calls. The same alignment knowledge also strengthens defenses: perception-aligned identity models can flag voice clones that metadata-based verification would accept. The benchmark measures model-human alignment at the evaluation level and does not release clone-generation recipes or adversarial training targets.",
+  "rai:dataLimitations": [
+    "English-speaking listener pool, US-dialect speakers. Cross-language perception is not measured.",
+    "100 speakers limits statistical power for some subgroup contrasts (20 speakers per sociophonetic group).",
+    "Studio-quality audio. In-the-wild conditions (noise, codec compression, telephony) are not represented.",
+    "The operational target is a population consensus, appropriate for ambiguous stimuli where any absolute identity label would itself be probabilistic."
+  ],
+  "rai:dataUseCases": [
+    "Benchmarking speaker embedding models against human voice-identity perception (Pearson r, R^2 against P(same)).",
+    "Human-aligned binary speaker verification (AUC, Platt-calibrated ECE against the listener majority vote).",
+    "Representational similarity analysis (RSA) between human and model representational dissimilarity matrices.",
+    "Real-to-synthetic distribution-shift evaluation (does a predictor fit on real-speech pairs still work on voice clones and morphs?).",
+    "Listener-conditioned identity modeling (using per-listener responses).",
+    "Calibration and uncertainty estimation in speaker verification."
+  ],
+  "rai:dataBiases": "Speakers are 100 US celebrities; the dataset over-represents US-dialect, professionally-recorded speech. Listeners are English-speaking adult crowdworkers. Generalization to non-English speakers, listeners, or in-the-wild audio conditions is not measured. The 5 sociophonetic groups carry 20 speakers each, limiting subgroup statistical power."
+}

data/speakers.csv

@@ -0,0 +1,101 @@
+id,name,group,gender,age
+M01,Vinny Guadagnino,1,1,1
+M02,Michael Gandolfini,1,1,1
+M03,Robert Iler,1,1,1
+M04,Sal Valentinetti,1,1,1
+M05,John Rondi Jr.,1,1,1
+M06,Oliver Anthony,2,1,1
+M07,Morgan Wallen,2,1,1
+M08,Luke Combs,2,1,1
+M09,Riley Green,2,1,1
+M10,Parker Mccollum,2,1,1
+M11,Kendrick Lamar,3,1,1
+M12,Tyler the Creator,3,1,1
+M13,J Cole,3,1,1
+M14,Michael B. Jordan,3,1,1
+M15,Lil Durk,3,1,1
+M16,Marcello Hernández,4,1,1
+M17,Pitbull,4,1,1
+M18,Maluma,4,1,1
+M19,Eric Hosmer,4,1,1
+M20,Robby Ramos,4,1,1
+M21,Jimmy O. Yang,5,1,1
+M22,Brandon Soo Hoo,5,1,1
+M23,Lawrence Kao,5,1,1
+M24,Ludi Lin,5,1,1
+M25,Andrew Fung,5,1,1
+M26,Michael Imperioli,1,1,2
+M27,Robert De Niro,1,1,2
+M28,Al Pacino,1,1,2
+M29,Harvey Keitel,1,1,2
+M30,Danny DeVito,1,1,2
+M31,Matthew McConaughey,2,1,2
+M32,Billy Bob Thornton,2,1,2
+M33,Jeff Foxworthy,2,1,2
+M34,Trace Adkins,2,1,2
+M35,Ricky Skaggs,2,1,2
+M36,Snoop Dogg,3,1,2
+M37,Ice Cube,3,1,2
+M38,Dr. Dre,3,1,2
+M39,Dave Chappelle,3,1,2
+M40,Jamie Foxx,3,1,2
+M41,Andy García,4,1,2
+M42,Steven Bauer,4,1,2
+M43,Ricky Martin,4,1,2
+M44,Marc Anthony,4,1,2
+M45,Eugenio Derbez,4,1,2
+M46,Tzi Ma,5,1,2
+M47,Hiroyuki Sanada,5,1,2
+M48,George Takei,5,1,2
+M49,BD Wong,5,1,2
+M50,James Hong,5,1,2
+F01,Cardi B,1,2,1
+F02,Lady Gaga,1,2,1
+F03,Julia Fox,1,2,1
+F04,Ice Spice,1,2,1
+F05,Angelina Pivarnick,1,2,1
+F06,Lainey Wilson,2,2,1
+F07,RaeLynn,2,2,1
+F08,Kacey Musgraves,2,2,1
+F09,Miley Cyrus,2,2,1
+F10,Kelsea Ballerini,2,2,1
+F11,Megan Thee Stallion,3,2,1
+F12,JT,3,2,1
+F13,Latto,3,2,1
+F14,Keke Palmer,3,2,1
+F15,Lizzo,3,2,1
+F16,Ana de Armas,4,2,1
+F17,Camila Cabello,4,2,1
+F18,Eiza González,4,2,1
+F19,Anitta,4,2,1
+F20,Lele Pons,4,2,1
+F21,Chloe Bennet,5,2,1
+F22,Stephanie Hsu,5,2,1
+F23,Kelly Marie Tran,5,2,1
+F24,Brenda Song,5,2,1
+F25,Ashley Park,5,2,1
+F26,Lorraine Bracco,1,2,2
+F27,Debi Mazar,1,2,2
+F28,Susie Essman,1,2,2
+F29,Edie Falco,1,2,2
+F30,Drea de Matteo,1,2,2
+F31,Dolly Parton,2,2,2
+F32,Reba McEntire,2,2,2
+F33,Sissy Spacek,2,2,2
+F34,Naomi Judd,2,2,2
+F35,Tanya Tucker,2,2,2
+F36,Mary J. Blige,3,2,2
+F37,Queen Latifah,3,2,2
+F38,Missy Elliott,3,2,2
+F39,Taraji P. Henson,3,2,2
+F40,Mo’Nique,3,2,2
+F41,Salma Hayek,4,2,2
+F42,Gloria Estefan,4,2,2
+F43,Jennifer Lopez,4,2,2
+F44,Rita Moreno,4,2,2
+F45,Kate del Castillo,4,2,2
+F46,Ming-Na Wen,5,2,2
+F47,Lucy Liu,5,2,2
+F48,Margaret Cho,5,2,2
+F49,Michelle Yeoh,5,2,2
+F50,Rosalind Chao,5,2,2

docs/annotation_protocol.md

@@ -0,0 +1,54 @@
+# Annotation protocol
+
+This document describes how the 124,876 listener judgments in VIPBench were collected.
+
+## Recruitment
+
+Listeners were recruited via the **Centaur AI** crowdsourcing platform (`https://centaur.ai`) under an Institutional Review Board (IRB) approved research protocol. The pool was restricted to English-speaking adults; 1,290 participants completed at least one trial. Compensation followed the platform's standard rate, which meets minimum-wage requirements in the country of data collection.
+
+## Consent
+
+Participants reviewed and accepted the platform's standard consent text before beginning the study. The IRB-approved protocol covered (a) the use of publicly available celebrity recordings as source audio, (b) the generation of AI voice clones from those recordings, and (c) the collection of binary same/different identity judgments.
+
+## Stimulus presentation
+
+Each trial presented a single audio clip in three parts:
+
+1. A reference recording of one celebrity speaker (~6 seconds).
+2. One second of silence followed by a short beep (separator).
+3. A comparison clip (real, AI cloned, or morphed; ~6 seconds).
+
+Stimuli are 16 kHz mono WAV. The combined stimulus is the file in `data/audio/`'s comparison directory together with the matching reference. The pre-concatenated trial-format clips used in the human study are not redistributed; users can recreate them by concatenating reference + 1 s silence + beep + comparison.
+
+## Response interface
+
+For each trial, listeners answered:
+
+1. **Identity judgment** (binary, required): "Are these two clips from the same speaker?" with response options "Same" or "Different".
+2. **Speaker recognition** (categorical, optional probe, recorded as `know_speaker`): which of four within-group celebrities (or "I don't know") they recognized in the reference clip.
+
+Response (1) defines the primary annotation; response (2) is used to filter trials where the listener recognized the reference (78.4% of all judgments come from unfamiliar-listener trials).
+
+## Coverage
+
+- Each pair received at least 8 judgments. Median 10, range 8 to 92.
+- Real-speech pairs received more coverage on average than synthetic pairs, giving tighter consensus estimates on the real-speech reference distribution.
+- Stimulus presentation order is randomized within participant.
+
+## Attention checks
+
+Embedded probes flagged inattentive responses; per-listener qualification flags are computed from these probes. The full release includes all responses; downstream users can apply the qualification filter via the columns in `data/participant_responses.csv`.
+
+## Demographics collected
+
+Per listener: pseudonymized integer identifier, age band, gender, and a binary first-language flag. No personally identifying information is included. Listener IDs are tied to no external account or platform user.
+
+## P(same) computation
+
+For each pair, `P(same) = same_vote / num_response` where `same_vote` counts listeners who answered "Same" and `num_response` is the total number of judgments on that pair. The Spearman-Brown corrected split-half reliability of `P(same)` over the 1,290-participant pool is `rho_SB = 0.705`, which bounds any model's correlation against the observed target.
+
+## Reproducing the listening study
+
+Researchers wishing to extend the study (e.g. with non-English listeners) can reuse this protocol. The trial-format audio files (reference + silence + beep + comparison) are deterministically reconstructable from `data/audio/`. Per-pair stimulus IDs in `data/stimuli.csv` allow exact replication of trial assignment.
+
+For the IRB-approval scope and permissible extensions, contact the dataset maintainers (camera-ready will list contact information).

docs/data_dictionary.md

@@ -0,0 +1,110 @@
+# Data dictionary
+
+Column-level schemas for every CSV in `data/` and `samples/`.
+
+## `speakers.csv` (100 rows)
+
+| Column | Type | Description | Valid values |
+|---|---|---|---|
+| `id` | text | Speaker identifier. Joins to `stimuli.reference` and to `<id>R` keys in embedding files. | F01-F50, M01-M50 |
+| `name` | text | Speaker name (public figure). | celebrity name |
+| `group` | int | Sociophonetic group code. | 1-5 (see below) |
+| `gender` | int | Speaker gender. | 1=male, 2=female |
+| `age` | int | Speaker age bracket. | 1=younger, 2=older |
+
+### Sociophonetic group mapping
+
+The 5 sociophonetic groups partition the 100 speakers into 20 speakers each. The integer-to-group mapping is documented by the dataset authors. The benchmark protocol uses `group` only as a stratification variable for fairness analyses; downstream users reproducing fairness analyses can rely on the integer codes directly.
+
+## `stimuli.csv` (9,800 rows)
+
+| Column | Type | Description |
+|---|---|---|
+| `id` | text | Stimulus identifier (matches the comparison-audio basename). Key into embedding files. |
+| `stimuli_type` | int | Stimulus type, 1-6. See `stimulus_types.md`. |
+| `reference` | text | Reference speaker ID. Joins to `speakers.id`. |
+| `comparison` | text | Comparison speaker ID for non-Type-6 pairs (NaN for Type 6). |
+| `voice_clone` | int | 1 if comparison clip is an AI voice clone, 0 otherwise. |
+| `correct_answer` | int | Metadata same/different label. 1=same speaker by metadata, 0=different. |
+| `scale` | int | For Type 6 morphs, interpolation level in [0, 100]. 100 for non-morph pairs. |
+| `num_response` | int | Number of listener judgments on this pair. |
+| `same_vote` | int | Listeners who answered "same speaker". |
+| `diff_vote` | int | Listeners who answered "different speaker". |
+| `correct_vote` | int | Listeners whose answer matches the metadata label. |
+| `incorrect_vote` | int | Listeners whose answer disagrees with the metadata label. |
+| `accuracy` | float | `correct_vote / num_response`. |
+| `group` | int | Reference speaker's sociophonetic group. |
+| `gender` | int | Reference speaker's gender. |
+| `age` | int | Reference speaker's age bracket. |
+
+`P(same)` is computed downstream as `same_vote / num_response`.
+
+## `participant_responses.csv` (124,876 rows)
+
+| Column | Type | Description |
+|---|---|---|
+| `user_id` | int | Pseudonymized listener identifier. Tied to no external account. |
+| `stimuli_id` | text | Stimulus identifier. Joins to `stimuli.id`. |
+| `stimuli_type` | int | Stimulus type, 1-6. |
+| `answer` | int | Listener's binary judgment. 1=same speaker, 0=different. |
+| `correct` | int | 1 if `answer` matches `correct_answer` in `stimuli.csv`, 0 otherwise. |
+| `know_speaker` | int | Listener-recognition probe. 1 if listener identified the reference speaker, 0 otherwise. May be missing for early-trial responses. |
+| `age` | float | Listener age band (categorical, encoded as float). |
+| `gender` | float | Listener gender. |
+| `first_language` | float | Listener first-language flag. 0=English first, 1=other. |
+| `num_stimuli_seen` | float | Cumulative stimulus count for this listener at the time of the response. |
+
+## `stimuli_interpol.csv` (8,100 rows)
+
+Per-stimulus metadata for Type 6 morphs.
+
+| Column | Type | Description |
+|---|---|---|
+| `id` | text | Stimulus identifier. Joins to `stimuli.id`. |
+| `source` | text | Source speaker A (one endpoint of the morph trajectory). |
+| `target` | text | Source speaker B (other endpoint). |
+| `scale` | int | Interpolation level in [0, 100]. |
+
+(Other columns may be present and are described in their headers; the four above are the schema-stable subset used by the analysis notebook.)
+
+## Embedding `.npz` files
+
+Each file in `data/embeddings/` is a key-value store:
+
+- **Keys** (`.files` attribute): audio basenames without `.wav`. 9,900 keys total: 100 references like `M01R`, `F03R` plus 9,800 comparisons like `1_F01`, `4_M12_M15B`, `6_F03_F09_50`.
+- **Values**: 1-D `np.float32` arrays of shape `(embedding_dim,)`. Dim depends on the model (see `model_table.md`).
+
+Per-layer files (`data/embeddings/layers/`) use the same keys; values have shape `(num_layers, embedding_dim)` (mean-pooled across time per layer).
+
+## How to pair reference and comparison
+
+Every voice pair in VIPBench is one row of `stimuli.csv`. The pairing rule is:
+
+| Asset | Reference clip | Comparison clip |
+|---|---|---|
+| CSV column | `reference` (e.g., `M01`) | `id` (e.g., `1_M01`, `4_M12_M15B`) |
+| Audio file | `data/audio/reference/{reference}R.wav` | `data/audio/comparison/{id}.wav` |
+| Embedding key | `{reference}R` (e.g., `M01R`) | `{id}` (e.g., `1_M01`) |
+
+Equivalently: the reference clip's basename is the speaker ID with `R` appended; the comparison clip's basename is exactly the stimulus `id`. The 100 reference embeddings (`*R`) and 9,800 comparison embeddings (`id`) together make the 9,900 keys present in every `.npz`.
+
+Reference example: scoring a model against `P(same)`.
+
+```python
+import numpy as np, pandas as pd
+from sklearn.metrics.pairwise import cosine_similarity
+
+stim = pd.read_csv('data/stimuli.csv')
+emb  = dict(np.load('data/embeddings/ecapa_tdnn.npz'))
+
+def cos_pair(row):
+    ref = emb[f'{row.reference}R'].reshape(1, -1)
+    cmp = emb[row.id].reshape(1, -1)
+    return cosine_similarity(ref, cmp)[0, 0]
+
+stim['cos']    = stim.apply(cos_pair, axis=1)
+stim['p_same'] = stim['same_vote'] / stim['num_response']
+print(stim[['cos', 'p_same']].corr())
+```
+
+The same pattern (with `librosa.load(...)` instead of dictionary lookup) loads the corresponding audio.

docs/model_table.md

@@ -0,0 +1,43 @@
+# 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 + Fisher + SWB | 192 | `nvidia/speakerverification_en_titanet_large` | `extract_titanet.py` |
+| resemblyzer | GE2E loss, 3-layer LSTM | LibriSpeech / VoxCeleb (mixed) | 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.

docs/reproduction.md

@@ -0,0 +1,108 @@
+# Reproduction
+
+End-to-end reproduction of the figures and tables in the accompanying NeurIPS 2026 paper.
+
+## Environment
+
+Tested on Linux x86_64 with Python 3.10 and 3.11. WSL2 also works.
+
+```bash
+python3 -m venv .venv
+source .venv/bin/activate
+pip install -r requirements.txt
+```
+
+GPU is optional. Whisper extraction is faster on GPU; everything else runs comfortably on CPU.
+
+## Quick reproduction (analysis only)
+
+This uses the embeddings already shipped in `data/embeddings/` and runs the analysis notebook end-to-end. **~10 minutes** on a laptop.
+
+```bash
+cd code
+bash reproduce.sh
+```
+
+The notebook is rendered to `code/benchmark_analysis_executed.ipynb`. Figures are written to `code/manuscript/figures/`.
+
+## Full reproduction (re-extract embeddings)
+
+This re-extracts embeddings from the audio in `data/audio/` for all 10 models. **~24 CPU-hours plus ~1 GPU-hour for Whisper.**
+
+```bash
+cd code
+bash reproduce.sh --extract
+```
+
+Or run individual extractors:
+
+```bash
+cd code
+python3 extract_ecapa_tdnn.py
+python3 extract_whisper.py
+# ...
+```
+
+Each script writes `<model>.npz` to `data/embeddings/`.
+
+## Path resolution
+
+The release directory is the implicit root. Override with `VIPBENCH_ROOT`:
+
+```bash
+VIPBENCH_ROOT=/path/to/vipbench_release bash code/reproduce.sh
+```
+
+The notebook auto-resolves paths via `Path.cwd().parent` when run from `code/`.
+
+## Caching
+
+The notebook caches expensive computations under `code/cache/`:
+
+- Cosine-similarity matrices per model
+- Bootstrap confidence intervals
+- Best-layer cosine similarities for SSL models
+- Mahalanobis-metric solutions
+
+Delete `code/cache/` to force recompute. The cache is keyed by content hash of inputs; changing the embeddings or CSVs invalidates the relevant entries automatically.
+
+## Outputs
+
+After successful execution, expect:
+
+- `code/benchmark_analysis_executed.ipynb`: rendered notebook
+- `code/manuscript/figures/*.png`: ~12 figures
+- `code/manuscript/tables/*.csv` (if generated by notebook): per-model summary tables
+
+## Verifying the reproduction
+
+The grand summary table (notebook cell near the end) should report:
+
+- resemblyzer: Pearson r ~0.656, R^2 ~0.430, AUC ~0.872
+- ECAPA-TDNN: Pearson r ~0.646
+- Whisper (final layer): Pearson r ~0.226
+- Spearman-Brown noise ceiling: rho_SB = 0.705
+
+Numbers should match the paper to 3 decimal places (small variation from random-seed effects in bootstrap CIs is expected).
+
+## Troubleshooting
+
+### NeMo / TitaNet install fails
+
+`nemo_toolkit[asr]` is heavy (~5 GB). If you don't need TitaNet, comment out its line in `requirements.txt` and skip `extract_titanet.py`. The notebook will print `SKIPPED titanet: no .npz found` and continue.
+
+### Hugging Face checkpoint download fails
+
+The extraction scripts download checkpoints on first run. Verify network access to `huggingface.co` and run `huggingface-cli login` if you've hit rate limits. Set `HF_HOME` to point to a persistent cache directory if you re-run frequently.
+
+### librosa / numba conflict
+
+If `pip install librosa` complains about numba, pin `numba<0.59` and re-install.
+
+### "No .npz found"
+
+The notebook logs `SKIPPED <model>: no .npz at <path>` when an embedding bundle is missing. The remaining models still produce results; the grand summary will simply omit the missing rows.
+
+### Out-of-memory during SSL layer extraction
+
+`extract_ssl_layers.py` keeps all layer activations in memory per model. For machines with less than 16 GB RAM, run one model at a time and clear caches between runs.

docs/stimulus_types.md

@@ -0,0 +1,46 @@
+# Stimulus types
+
+VIPBench contains 9,800 voice pairs in **6 stimulus types**, designed to span the perceptual identity landscape from "obviously same" to "obviously different" with controlled ambiguous regions in between.
+
+## The six types
+
+| Type | Description | Metadata label | Pair count | P(same) shape |
+|---|---|---|---|---|
+| 1 | Same recording (reference compared with itself, segmented differently) | Same | 100 | concentrated near 1.0 |
+| 2 | Same speaker, different recording | Same | 500 | high but spread |
+| 3 | Same speaker, AI voice clone | Same | 500 | spreads across full range (clones with metadata-same label) |
+| 4 | Different speakers, real recordings | Different | 500 | concentrated near 0.0 |
+| 5 | Different speakers, AI voice clones | Different | 100 | concentrated near 0.0 |
+| 6 | Continuously morphed voices | (no clean metadata label) | 8,100 | sweeps full range across the morph trajectory |
+
+Total: **9,800 pairs**, of which 6,100 carry a clean metadata same/different label (Types 1-5) and 8,100 are morph trajectories.
+
+Note: Type 6 pair count of 8,100 reflects 100 morph-trajectory cells, each sampled at 81 levels of interpolation; see `data/stimuli_interpol.csv` for trajectory metadata.
+
+## Naming convention
+
+Stimulus IDs in `data/stimuli.csv` and the audio filenames in `data/audio/comparison/` follow the pattern:
+
+- **Types 1-5:** `<type>_<reference_speaker>` for type-1 same-recording pairs; `<type>_<reference_speaker><variant>` where the comparison clip varies across A-E for types 2-5.
+  - Examples: `1_M01.wav` (Type 1, M01), `2_M01B.wav` (Type 2, M01 with variant B), `4_F03_F09B.wav` (Type 4, reference F03 paired with F09 variant B).
+- **Type 6 morphs:** `6_<source>_<target>_<scale>.wav` where `<scale>` is an integer in [0, 100] indicating the interpolation point.
+
+The stimulus ID matches the comparison-audio basename (without `.wav`) and is the key into the embedding `.npz` files.
+
+## Voice cloning
+
+Voice clones (Types 3 and 5) were generated with a state-of-the-art TTS system from a short reference clip per speaker. The clone shares the metadata identity of the source speaker by construction; whether listeners hear the clone as that speaker is the per-pair question that the benchmark measures via `P(same)`.
+
+## Voice morphing
+
+Type 6 pairs are continuously morphed by interpolating a voice-conversion latent between two source speakers at 81 levels per pair. Morphs have no clean metadata speaker label: at scale 0 the audio matches one speaker, at scale 100 the other, and intermediate scales sweep a perceptual continuum. This is the largest category in the dataset (8,100 of 9,800 pairs) and is designed to probe identity perception at fine resolution. The corresponding trajectory metadata (source speakers, scale value) is in `data/stimuli_interpol.csv`.
+
+## Why this design
+
+The six types reflect three orthogonal axes:
+
+1. **Metadata identity** (same vs different speaker): Types 1-2-3 vs 4-5; Type 6 sweeps.
+2. **Synthesis** (real vs AI-generated): Types 1-2-4 vs 3-5; Type 6 is morphed.
+3. **Ambiguity** (concentrated vs spread `P(same)`): Types 1, 4, 5 are concentrated; Types 3 and 6 sweep, exposing where listener perception diverges from the metadata label.
+
+Types 3 and 6 are where perceptual and metadata identity most often disagree, making them the centerpiece of the benchmark's perception-vs-metadata contrast.

requirements.txt

@@ -0,0 +1,34 @@
+# Core scientific stack
+numpy>=1.24,<2.0
+pandas>=2.0,<3.0
+scipy>=1.10
+scikit-learn>=1.3
+matplotlib>=3.7
+seaborn>=0.12
+statsmodels>=0.14
+tqdm>=4.65
+
+# Audio I/O
+librosa>=0.10
+soundfile>=0.12
+
+# Notebook (analysis)
+jupyter>=1.0
+notebook>=7.0
+nbconvert>=7.0
+
+# Speaker / speech representation models
+torch>=2.0,<3.0
+torchaudio>=2.0,<3.0
+transformers>=4.35,<5.0       # wav2vec2, HuBERT, WavLM, XLS-R
+speechbrain>=1.0              # ECAPA-TDNN
+openai-whisper>=20231117      # Whisper
+resemblyzer>=0.1.4            # GE2E (resemblyzer)
+
+# Optional, heavy install (uncomment to enable TitaNet extraction):
+# nemo_toolkit[asr]>=1.21.0   # NVIDIA NeMo (TitaNet)
+
+# Notes:
+# - x-vector and RawNet3 use the speechbrain / espnet ecosystem; specific
+#   checkpoint download is handled by the extraction scripts.
+# - The release ships pre-extracted embeddings, so re-running extraction is optional.

samples/README.md

@@ -0,0 +1,71 @@
+# VIPBench reviewer-friendly sample subset
+
+This directory holds a 5-speaker slice of the full release so reviewers can inspect the dataset without downloading the full ~5.8 GB bundle. Bundle size: **~115 MB** (audio + subsetted embeddings).
+
+The format is identical to the full release; analyses that work on `data/` work on `samples/` by changing one path.
+
+## What's included
+
+| Item | Count |
+|---|---|
+| Speakers | 5 (M01, F06, M11, F16, M21; one per sociophonetic group, 3M+2F) |
+| Reference audio (R.wav) | 5 |
+| Comparison audio | 490 (98 per speaker, all stimulus types) |
+| Listener judgments | 6,401 |
+| Pre-extracted embeddings | 10 models, subsetted to 495 keys each |
+| Per-layer SSL embeddings | 5 models, subsetted to 495 keys each |
+| Stimulus types covered | All 6 |
+
+## Layout
+
+```
+samples/
+  README.md                        # this file
+  speakers.csv                     # 5 rows (subset of data/speakers.csv)
+  stimuli.csv                      # 490 rows (subset of data/stimuli.csv)
+  participant_responses.csv        # 6,401 rows (subset of data/participant_responses.csv)
+  audio/
+    reference/                     # 5 *R.wav (symlinks to ../../../exp_2)
+    comparison/                    # 490 *.wav (symlinks to ../../../output)
+  embeddings/
+    rawnet3.npz, ecapa_tdnn.npz, ... (10 models)
+    layers/wav2vec2.npz, ... (5 SSL models)
+```
+
+## Quickstart
+
+```python
+import numpy as np, pandas as pd
+from sklearn.metrics.pairwise import cosine_similarity
+
+stim = pd.read_csv('samples/stimuli.csv')
+emb = dict(np.load('samples/embeddings/ecapa_tdnn.npz'))
+
+cos = []
+for _, row in stim.iterrows():
+    ref, cmp = emb[f'{row.reference}R'], emb[row.id]
+    cos.append(cosine_similarity([ref], [cmp])[0][0])
+stim['cos_ecapa'] = cos
+
+# P(same) target
+stim['p_same'] = stim['same_vote'] / stim['num_response']
+
+# Pearson r
+print(stim[['cos_ecapa', 'p_same']].corr())
+```
+
+## Speaker subset
+
+| ID | Group | Gender | Age bracket |
+|---|---|---|---|
+| M01 | 1 | M | 1 |
+| F06 | 2 | F | 1 |
+| M11 | 3 | M | 1 |
+| F16 | 4 | F | 1 |
+| M21 | 5 | M | 1 |
+
+The `group` column is an integer code for the speaker's sociophonetic group (5 groups, 20 speakers each). See `docs/data_dictionary.md` for the integer-to-group mapping.
+
+## License
+
+CC-BY-NC 4.0 for audio + judgments + embeddings; same as the full release. See `../LICENSE`.

samples/audio/reference/F06R.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2351910b9a2f3b0aec4595a91fdde7883414c66ce98d4c53acf9f99731db6b83
+size 454632

samples/audio/reference/F16R.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b4d8deb6c7caa938a5a0b728de8b0228a1697f021bd1d229e466af9dd6abb41b
+size 347568

samples/audio/reference/M11R.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:12f07d941f5c7a10ad2c9da3308cd7a3a64d3e2dccdf5d1dc5705a474487940c
+size 385600

samples/embeddings/ecapa_tdnn.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c55b7bd8599f70ddbab8f1f99e023e2a158f29b3242e2db5dcc8cf99e148030a
+size 484146

samples/embeddings/hubert.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b5d4d8939a39420082ee683acfb76e53acaa67e4884aea2d1d19b338b97ee161
+size 1541888

samples/embeddings/rawnet3.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e926bcf360cca11b6e614a9bd13e629e59af2e82fc9ee003a2bcbf66e64a1421
+size 484498

samples/embeddings/resemblyzer.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2e3f87ae717faf379acbae581a985f327e26dab8907a4bbc77f5809398244e72
+size 455399

samples/embeddings/titanet.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:93c98fa22ab683bd33dfe991a65d05947eeec893c9c8a97847ff68acbddbc01d
+size 483731

samples/embeddings/wav2vec2.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8338ee10b7b318f380b547bd024d52734771fe692a653f418389ae64828a265b
+size 1536096

samples/embeddings/wavlm.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8002e8e8d85a6b4268104ea8fe151c6a40ff8e0d0576a1dd02efb3bb630d122b
+size 1540400

samples/embeddings/whisper.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:acba595acca373dc2e014b7415d9d5c4b107c4812a7a1c0db882bb865ab03be7
+size 1066617

samples/embeddings/xlsr.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cdca0334789c840c1a8efe0dc8874b19803b259e14b1591e7fe867223a678316
+size 2013391

samples/embeddings/xvector.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8c5cf969e44d2db5f31a0a31188a1d7da561aa83bf4e6059f572dae2e39d1ba4
+size 1055035

samples/speakers.csv

@@ -0,0 +1,6 @@
+id,name,group,gender,age
+M01,Vinny Guadagnino,1,1,1
+M11,Kendrick Lamar,3,1,1
+M21,Jimmy O. Yang,5,1,1
+F06,Lainey Wilson,2,2,1
+F16,Ana de Armas,4,2,1

samples/stimuli.csv

@@ -0,0 +1,491 @@
+id,stimuli_type,reference,comparison,voice_clone,correct_answer,scale,num_response,same_vote,diff_vote,correct_vote,incorrect_vote,accuracy,group,gender,age
+1_M01,1,M01,M01,0,1,100,11,11,0,11,0,1.0,1,1,1
+2_M01B,2,M01,M01,0,1,100,75,34,41,34,41,0.4533333333333333,1,1,1
+2_M01C,2,M01,M01,0,1,100,35,31,4,31,4,0.8857142857142857,1,1,1
+2_M01D,2,M01,M01,0,1,100,35,14,21,14,21,0.4,1,1,1
+2_M01E,2,M01,M01,0,1,100,32,12,20,12,20,0.375,1,1,1
+3_M01B,3,M01,M01,1,1,100,11,8,3,8,3,0.7272727272727273,1,1,1
+3_M01C,3,M01,M01,1,1,100,11,10,1,10,1,0.9090909090909092,1,1,1
+3_M01D,3,M01,M01,1,1,100,10,5,5,5,5,0.5,1,1,1
+3_M01E,3,M01,M01,1,1,100,10,7,3,7,3,0.7,1,1,1
+4_M01_M05D,4,M01,M05,0,0,100,37,5,32,32,5,0.8648648648648649,1,1,1
+4_M01_M04D,4,M01,M04,0,0,100,35,2,33,33,2,0.9428571428571428,1,1,1
+4_M01_M02C,4,M01,M02,0,0,100,40,8,32,32,8,0.8,1,1,1
+4_M01_M03E,4,M01,M03,0,0,100,37,5,32,32,5,0.8648648648648649,1,1,1
+5_M01_M05D,5,M01,M05,1,0,100,10,2,8,8,2,0.8,1,1,1
+5_M01_M04D,5,M01,M04,1,0,100,10,4,6,6,4,0.6,1,1,1
+5_M01_M02C,5,M01,M02,1,0,100,10,4,6,6,4,0.6,1,1,1
+5_M01_M03E,5,M01,M03,1,0,100,10,4,6,6,4,0.6,1,1,1
+6_M01A_M02A_057,6,M01,M02,1,0,57,10,10,0,0,10,0.0,1,1,1
+6_M01A_M05A_098,6,M01,M05,1,0,98,10,4,6,6,4,0.6,1,1,1
+6_M01A_M03A_054,6,M01,M03,1,0,54,10,3,7,7,3,0.7,1,1,1
+6_M01A_M05A_042,6,M01,M05,1,0,42,10,5,5,5,5,0.5,1,1,1
+6_M01A_M03A_037,6,M01,M03,1,0,37,10,3,7,7,3,0.7,1,1,1
+6_M01A_M02A_073,6,M01,M02,1,0,73,10,8,2,2,8,0.2,1,1,1
+6_M01A_M03A_009,6,M01,M03,1,0,9,10,2,8,8,2,0.8,1,1,1
+6_M01A_M02A_053,6,M01,M02,1,0,53,10,7,3,3,7,0.3,1,1,1
+6_M01A_M04A_067,6,M01,M04,1,0,67,10,7,3,3,7,0.3,1,1,1
+6_M01A_M02A_014,6,M01,M02,1,0,14,10,2,8,8,2,0.8,1,1,1
+6_M01A_M03A_053,6,M01,M03,1,0,53,10,2,8,8,2,0.8,1,1,1
+6_M01A_M02A_010,6,M01,M02,1,0,10,10,4,6,6,4,0.6,1,1,1
+6_M01A_M01A_100,6,M01,M01,1,0,100,10,6,4,4,6,0.4,1,1,1
+6_M01A_M03A_072,6,M01,M03,1,0,72,10,8,2,2,8,0.2,1,1,1
+6_M01A_M02A_090,6,M01,M02,1,0,90,10,8,2,2,8,0.2,1,1,1
+6_M01A_M04A_052,6,M01,M04,1,0,52,10,1,9,9,1,0.9,1,1,1
+6_M01A_M05A_084,6,M01,M05,1,0,84,10,9,1,1,9,0.1,1,1,1
+6_M01A_M04A_045,6,M01,M04,1,0,45,10,1,9,9,1,0.9,1,1,1
+6_M01A_M02A_009,6,M01,M02,1,0,9,10,2,8,8,2,0.8,1,1,1
+6_M01A_M05A_074,6,M01,M05,1,0,74,10,7,3,3,7,0.3,1,1,1
+6_M01A_M05A_035,6,M01,M05,1,0,35,10,2,8,8,2,0.8,1,1,1
+6_M01A_M04A_007,6,M01,M04,1,0,7,10,1,9,9,1,0.9,1,1,1
+6_M01A_M05A_056,6,M01,M05,1,0,56,10,2,8,8,2,0.8,1,1,1
+6_M01A_M05A_014,6,M01,M05,1,0,14,10,2,8,8,2,0.8,1,1,1
+6_M01A_M04A_026,6,M01,M04,1,0,26,10,3,7,7,3,0.7,1,1,1
+6_M01A_M04A_024,6,M01,M04,1,0,24,10,2,8,8,2,0.8,1,1,1
+6_M01A_M04A_022,6,M01,M04,1,0,22,10,1,9,9,1,0.9,1,1,1
+6_M01A_M03A_095,6,M01,M03,1,0,95,10,9,1,1,9,0.1,1,1,1
+6_M01A_M02A_003,6,M01,M02,1,0,3,10,2,8,8,2,0.8,1,1,1
+6_M01A_M03A_059,6,M01,M03,1,0,59,10,3,7,7,3,0.7,1,1,1
+6_M01A_M04A_032,6,M01,M04,1,0,32,10,1,9,9,1,0.9,1,1,1
+6_M01A_M05A_053,6,M01,M05,1,0,53,10,5,5,5,5,0.5,1,1,1
+6_M01A_M02A_024,6,M01,M02,1,0,24,10,3,7,7,3,0.7,1,1,1
+6_M01A_M03A_083,6,M01,M03,1,0,83,10,7,3,3,7,0.3,1,1,1
+6_M01A_M03A_081,6,M01,M03,1,0,81,10,7,3,3,7,0.3,1,1,1
+6_M01A_M03A_079,6,M01,M03,1,0,79,10,4,6,6,4,0.6,1,1,1
+6_M01A_M04A_091,6,M01,M04,1,0,91,10,8,2,2,8,0.2,1,1,1
+6_M01A_M04A_062,6,M01,M04,1,0,62,10,6,4,4,6,0.4,1,1,1
+6_M01A_M05A_048,6,M01,M05,1,0,48,10,6,4,4,6,0.4,1,1,1
+6_M01A_M02A_011,6,M01,M02,1,0,11,10,3,7,7,3,0.7,1,1,1
+6_M01A_M05A_073,6,M01,M05,1,0,73,10,6,4,4,6,0.4,1,1,1
+6_M01B_M02B_019,6,M01,M02,1,0,19,10,2,8,8,2,0.8,1,1,1
+6_M01B_M04B_022,6,M01,M04,1,0,22,10,1,9,9,1,0.9,1,1,1
+6_M01B_M03B_053,6,M01,M03,1,0,53,10,2,8,8,2,0.8,1,1,1
+6_M01B_M03B_083,6,M01,M03,1,0,83,10,3,7,7,3,0.7,1,1,1
+6_M01B_M04B_007,6,M01,M04,1,0,7,10,1,9,9,1,0.9,1,1,1
+6_M01B_M05B_074,6,M01,M05,1,0,74,10,3,7,7,3,0.7,1,1,1
+6_M01B_M02B_003,6,M01,M02,1,0,3,10,5,5,5,5,0.5,1,1,1
+6_M01B_M03B_079,6,M01,M03,1,0,79,10,3,7,7,3,0.7,1,1,1
+6_M01B_M04B_091,6,M01,M04,1,0,91,10,5,5,5,5,0.5,1,1,1
+6_M01B_M03B_037,6,M01,M03,1,0,37,10,1,9,9,1,0.9,1,1,1
+6_M01B_M04B_026,6,M01,M04,1,0,26,10,2,8,8,2,0.8,1,1,1
+6_M01B_M04B_024,6,M01,M04,1,0,24,11,0,11,11,0,1.0,1,1,1
+6_M01B_M04B_067,6,M01,M04,1,0,67,10,2,8,8,2,0.8,1,1,1
+6_M01B_M02B_011,6,M01,M02,1,0,11,10,4,6,6,4,0.6,1,1,1
+6_M01B_M05B_035,6,M01,M05,1,0,35,10,3,7,7,3,0.7,1,1,1
+6_M01B_M03B_095,6,M01,M03,1,0,95,10,6,4,4,6,0.4,1,1,1
+6_M01B_M05B_084,6,M01,M05,1,0,84,10,3,7,7,3,0.7,1,1,1
+6_M01B_M05B_014,6,M01,M05,1,0,14,10,1,9,9,1,0.9,1,1,1
+6_M01B_M05B_056,6,M01,M05,1,0,56,10,3,7,7,3,0.7,1,1,1
+6_M01B_M05B_053,6,M01,M05,1,0,53,10,2,8,8,2,0.8,1,1,1
+6_M01B_M05B_073,6,M01,M05,1,0,73,10,3,7,7,3,0.7,1,1,1
+6_M01B_M05B_048,6,M01,M05,1,0,48,10,2,8,8,2,0.8,1,1,1
+6_M01B_M04B_062,6,M01,M04,1,0,62,10,3,7,7,3,0.7,1,1,1
+6_M01B_M04B_032,6,M01,M04,1,0,32,10,1,9,9,1,0.9,1,1,1
+6_M01B_M02B_009,6,M01,M02,1,0,9,10,6,4,4,6,0.4,1,1,1
+6_M01B_M02B_014,6,M01,M02,1,0,14,10,7,3,3,7,0.3,1,1,1
+6_M01B_M04B_052,6,M01,M04,1,0,52,10,1,9,9,1,0.9,1,1,1
+6_M01B_M02B_024,6,M01,M02,1,0,24,10,7,3,3,7,0.3,1,1,1
+6_M01B_M02B_073,6,M01,M02,1,0,73,11,4,7,7,4,0.6363636363636364,1,1,1
+6_M01B_M04B_045,6,M01,M04,1,0,45,10,1,9,9,1,0.9,1,1,1
+6_M01B_M03B_081,6,M01,M03,1,0,81,10,3,7,7,3,0.7,1,1,1
+6_M01B_M05B_042,6,M01,M05,1,0,42,10,1,9,9,1,0.9,1,1,1
+6_M01B_M03B_059,6,M01,M03,1,0,59,10,4,6,6,4,0.6,1,1,1
+6_M01B_M05B_098,6,M01,M05,1,0,98,10,4,6,6,4,0.6,1,1,1
+6_M01B_M03B_054,6,M01,M03,1,0,54,10,4,6,6,4,0.6,1,1,1
+6_M01B_M02B_053,6,M01,M02,1,0,53,10,5,5,5,5,0.5,1,1,1
+6_M01B_M03B_072,6,M01,M03,1,0,72,10,7,3,3,7,0.3,1,1,1
+6_M01B_M02B_057,6,M01,M02,1,0,57,10,4,6,6,4,0.6,1,1,1
+6_M01B_M03B_009,6,M01,M03,1,0,9,10,2,8,8,2,0.8,1,1,1
+6_M01B_M02B_010,6,M01,M02,1,0,10,10,5,5,5,5,0.5,1,1,1
+1_M11,1,M11,M11,0,1,100,35,35,0,35,0,1.0,3,1,1
+2_M11B,2,M11,M11,0,1,100,36,7,29,7,29,0.1944444444444444,3,1,1
+2_M11C,2,M11,M11,0,1,100,30,10,20,10,20,0.3333333333333333,3,1,1
+2_M11D,2,M11,M11,0,1,100,32,13,19,13,19,0.40625,3,1,1
+2_M11E,2,M11,M11,0,1,100,88,65,23,65,23,0.7386363636363636,3,1,1
+3_M11B,3,M11,M11,1,1,100,10,4,6,4,6,0.4,3,1,1
+3_M11C,3,M11,M11,1,1,100,10,6,4,6,4,0.6,3,1,1
+3_M11D,3,M11,M11,1,1,100,10,2,8,2,8,0.2,3,1,1
+3_M11E,3,M11,M11,1,1,100,10,4,6,4,6,0.4,3,1,1
+4_M11_M13C,4,M11,M13,0,0,100,36,5,31,31,5,0.8611111111111112,3,1,1
+4_M11_M12C,4,M11,M12,0,0,100,33,6,27,27,6,0.8181818181818182,3,1,1
+4_M11_M15E,4,M11,M15,0,0,100,40,23,17,17,23,0.425,3,1,1
+4_M11_M14C,4,M11,M14,0,0,100,30,4,26,26,4,0.8666666666666667,3,1,1
+5_M11_M13C,5,M11,M13,1,0,100,10,4,6,6,4,0.6,3,1,1
+5_M11_M12C,5,M11,M12,1,0,100,10,3,7,7,3,0.7,3,1,1
+5_M11_M15E,5,M11,M15,1,0,100,10,8,2,2,8,0.2,3,1,1
+5_M11_M14C,5,M11,M14,1,0,100,10,5,5,5,5,0.5,3,1,1
+6_M11A_M14A_052,6,M11,M14,1,0,52,10,0,10,10,0,1.0,3,1,1
+6_M11A_M12A_092,6,M11,M12,1,0,92,10,5,5,5,5,0.5,3,1,1
+6_M11A_M12A_005,6,M11,M12,1,0,5,10,0,10,10,0,1.0,3,1,1
+6_M11A_M13A_039,6,M11,M13,1,0,39,11,4,7,7,4,0.6363636363636364,3,1,1
+6_M11A_M13A_014,6,M11,M13,1,0,14,10,3,7,7,3,0.7,3,1,1
+6_M11A_M15A_045,6,M11,M15,1,0,45,10,2,8,8,2,0.8,3,1,1
+6_M11A_M15A_035,6,M11,M15,1,0,35,10,3,7,7,3,0.7,3,1,1
+6_M11A_M12A_041,6,M11,M12,1,0,41,11,5,6,6,5,0.5454545454545454,3,1,1
+6_M11A_M15A_065,6,M11,M15,1,0,65,10,4,6,6,4,0.6,3,1,1
+6_M11A_M13A_069,6,M11,M13,1,0,69,10,8,2,2,8,0.2,3,1,1
+6_M11A_M14A_021,6,M11,M14,1,0,21,10,1,9,9,1,0.9,3,1,1
+6_M11A_M15A_043,6,M11,M15,1,0,43,10,3,7,7,3,0.7,3,1,1
+6_M11A_M15A_008,6,M11,M15,1,0,8,10,1,9,9,1,0.9,3,1,1
+6_M11A_M12A_040,6,M11,M12,1,0,40,10,1,9,9,1,0.9,3,1,1
+6_M11A_M14A_034,6,M11,M14,1,0,34,10,1,9,9,1,0.9,3,1,1
+6_M11A_M15A_073,6,M11,M15,1,0,73,10,9,1,1,9,0.1,3,1,1
+6_M11A_M15A_031,6,M11,M15,1,0,31,10,1,9,9,1,0.9,3,1,1
+6_M11A_M13A_018,6,M11,M13,1,0,18,10,2,8,8,2,0.8,3,1,1
+6_M11A_M14A_011,6,M11,M14,1,0,11,10,0,10,10,0,1.0,3,1,1
+6_M11A_M13A_030,6,M11,M13,1,0,30,10,3,7,7,3,0.7,3,1,1
+6_M11A_M15A_042,6,M11,M15,1,0,42,10,3,7,7,3,0.7,3,1,1
+6_M11A_M14A_042,6,M11,M14,1,0,42,10,2,8,8,2,0.8,3,1,1
+6_M11A_M12A_050,6,M11,M12,1,0,50,10,2,8,8,2,0.8,3,1,1
+6_M11A_M15A_055,6,M11,M15,1,0,55,11,8,3,3,8,0.2727272727272727,3,1,1
+6_M11A_M14A_050,6,M11,M14,1,0,50,10,0,10,10,0,1.0,3,1,1
+6_M11A_M12A_086,6,M11,M12,1,0,86,10,6,4,4,6,0.4,3,1,1
+6_M11A_M12A_028,6,M11,M12,1,0,28,10,3,7,7,3,0.7,3,1,1
+6_M11A_M14A_059,6,M11,M14,1,0,59,10,4,6,6,4,0.6,3,1,1
+6_M11A_M13A_026,6,M11,M13,1,0,26,10,4,6,6,4,0.6,3,1,1
+6_M11A_M13A_085,6,M11,M13,1,0,85,10,7,3,3,7,0.3,3,1,1
+6_M11A_M14A_057,6,M11,M14,1,0,57,10,5,5,5,5,0.5,3,1,1
+6_M11A_M14A_066,6,M11,M14,1,0,66,10,6,4,4,6,0.4,3,1,1
+6_M11A_M11A_100,6,M11,M11,1,0,100,11,8,3,3,8,0.2727272727272727,3,1,1
+6_M11A_M12A_090,6,M11,M12,1,0,90,10,9,1,1,9,0.1,3,1,1
+6_M11A_M13A_033,6,M11,M13,1,0,33,10,6,4,4,6,0.4,3,1,1
+6_M11A_M12A_010,6,M11,M12,1,0,10,10,4,6,6,4,0.6,3,1,1
+6_M11A_M12A_031,6,M11,M12,1,0,31,10,4,6,6,4,0.6,3,1,1
+6_M11A_M13A_001,6,M11,M13,1,0,1,10,2,8,8,2,0.8,3,1,1
+6_M11A_M15A_039,6,M11,M15,1,0,39,10,1,9,9,1,0.9,3,1,1
+6_M11A_M13A_055,6,M11,M13,1,0,55,10,5,5,5,5,0.5,3,1,1
+6_M11A_M14A_041,6,M11,M14,1,0,41,11,1,10,10,1,0.9090909090909092,3,1,1
+6_M11B_M13B_055,6,M11,M13,1,0,55,10,0,10,10,0,1.0,3,1,1
+6_M11B_M12B_041,6,M11,M12,1,0,41,10,0,10,10,0,1.0,3,1,1
+6_M11B_M13B_039,6,M11,M13,1,0,39,10,4,6,6,4,0.6,3,1,1
+6_M11B_M14B_050,6,M11,M14,1,0,50,10,1,9,9,1,0.9,3,1,1
+6_M11B_M15B_035,6,M11,M15,1,0,35,10,2,8,8,2,0.8,3,1,1
+6_M11B_M15B_045,6,M11,M15,1,0,45,10,1,9,9,1,0.9,3,1,1
+6_M11B_M12B_092,6,M11,M12,1,0,92,11,2,9,9,2,0.8181818181818182,3,1,1
+6_M11B_M13B_014,6,M11,M13,1,0,14,10,3,7,7,3,0.7,3,1,1
+6_M11B_M15B_055,6,M11,M15,1,0,55,10,2,8,8,2,0.8,3,1,1
+6_M11B_M14B_034,6,M11,M14,1,0,34,10,2,8,8,2,0.8,3,1,1
+6_M11B_M14B_059,6,M11,M14,1,0,59,10,1,9,9,1,0.9,3,1,1
+6_M11B_M14B_052,6,M11,M14,1,0,52,10,1,9,9,1,0.9,3,1,1
+6_M11B_M15B_008,6,M11,M15,1,0,8,10,0,10,10,0,1.0,3,1,1
+6_M11B_M12B_005,6,M11,M12,1,0,5,10,2,8,8,2,0.8,3,1,1
+6_M11B_M12B_086,6,M11,M12,1,0,86,10,2,8,8,2,0.8,3,1,1
+6_M11B_M14B_011,6,M11,M14,1,0,11,10,1,9,9,1,0.9,3,1,1
+6_M11B_M15B_073,6,M11,M15,1,0,73,10,0,10,10,0,1.0,3,1,1
+6_M11B_M13B_069,6,M11,M13,1,0,69,10,2,8,8,2,0.8,3,1,1
+6_M11B_M12B_010,6,M11,M12,1,0,10,10,1,9,9,1,0.9,3,1,1
+6_M11B_M13B_018,6,M11,M13,1,0,18,10,2,8,8,2,0.8,3,1,1
+6_M11B_M14B_042,6,M11,M14,1,0,42,10,2,8,8,2,0.8,3,1,1
+6_M11B_M12B_050,6,M11,M12,1,0,50,10,2,8,8,2,0.8,3,1,1
+6_M11B_M13B_026,6,M11,M13,1,0,26,10,2,8,8,2,0.8,3,1,1
+6_M11B_M15B_039,6,M11,M15,1,0,39,10,1,9,9,1,0.9,3,1,1
+6_M11B_M12B_090,6,M11,M12,1,0,90,10,1,9,9,1,0.9,3,1,1
+6_M11B_M14B_041,6,M11,M14,1,0,41,10,2,8,8,2,0.8,3,1,1
+6_M11B_M12B_028,6,M11,M12,1,0,28,10,4,6,6,4,0.6,3,1,1
+6_M11B_M15B_031,6,M11,M15,1,0,31,10,1,9,9,1,0.9,3,1,1
+6_M11B_M13B_033,6,M11,M13,1,0,33,10,2,8,8,2,0.8,3,1,1
+6_M11B_M12B_031,6,M11,M12,1,0,31,10,0,10,10,0,1.0,3,1,1
+6_M11B_M15B_065,6,M11,M15,1,0,65,10,1,9,9,1,0.9,3,1,1
+6_M11B_M14B_057,6,M11,M14,1,0,57,10,2,8,8,2,0.8,3,1,1
+6_M11B_M14B_066,6,M11,M14,1,0,66,10,2,8,8,2,0.8,3,1,1
+6_M11B_M15B_043,6,M11,M15,1,0,43,10,2,8,8,2,0.8,3,1,1
+6_M11B_M15B_042,6,M11,M15,1,0,42,11,4,7,7,4,0.6363636363636364,3,1,1
+6_M11B_M14B_021,6,M11,M14,1,0,21,10,0,10,10,0,1.0,3,1,1
+6_M11B_M13B_001,6,M11,M13,1,0,1,10,3,7,7,3,0.7,3,1,1
+6_M11B_M13B_030,6,M11,M13,1,0,30,10,2,8,8,2,0.8,3,1,1
+6_M11B_M13B_085,6,M11,M13,1,0,85,10,1,9,9,1,0.9,3,1,1
+6_M11B_M12B_040,6,M11,M12,1,0,40,10,0,10,10,0,1.0,3,1,1
+1_M21,1,M21,M21,0,1,100,30,24,6,24,6,0.8,5,1,1
+2_M21B,2,M21,M21,0,1,100,71,52,19,52,19,0.7323943661971831,5,1,1
+2_M21C,2,M21,M21,0,1,100,36,7,29,7,29,0.1944444444444444,5,1,1
+2_M21D,2,M21,M21,0,1,100,36,21,15,21,15,0.5833333333333334,5,1,1
+2_M21E,2,M21,M21,0,1,100,28,26,2,26,2,0.9285714285714286,5,1,1
+3_M21B,3,M21,M21,1,1,100,10,6,4,6,4,0.6,5,1,1
+3_M21C,3,M21,M21,1,1,100,10,4,6,4,6,0.4,5,1,1
+3_M21D,3,M21,M21,1,1,100,10,1,9,1,9,0.1,5,1,1
+3_M21E,3,M21,M21,1,1,100,10,8,2,8,2,0.8,5,1,1
+4_M21_M22D,4,M21,M22,0,0,100,36,9,27,27,9,0.75,5,1,1
+4_M21_M24B,4,M21,M24,0,0,100,34,8,26,26,8,0.7647058823529411,5,1,1
+4_M21_M25B,4,M21,M25,0,0,100,70,33,37,37,33,0.5285714285714286,5,1,1
+4_M21_M23D,4,M21,M23,0,0,100,37,6,31,31,6,0.8378378378378378,5,1,1
+5_M21_M22D,5,M21,M22,1,0,100,10,2,8,8,2,0.8,5,1,1
+5_M21_M24B,5,M21,M24,1,0,100,10,2,8,8,2,0.8,5,1,1
+5_M21_M25B,5,M21,M25,1,0,100,10,5,5,5,5,0.5,5,1,1
+5_M21_M23D,5,M21,M23,1,0,100,10,2,8,8,2,0.8,5,1,1
+6_M21A_M21A_100,6,M21,M21,1,0,100,10,7,3,3,7,0.3,5,1,1
+6_M21A_M24A_061,6,M21,M24,1,0,61,10,7,3,3,7,0.3,5,1,1
+6_M21A_M24A_091,6,M21,M24,1,0,91,10,4,6,6,4,0.6,5,1,1
+6_M21A_M24A_033,6,M21,M24,1,0,33,10,1,9,9,1,0.9,5,1,1
+6_M21A_M22A_077,6,M21,M22,1,0,77,10,5,5,5,5,0.5,5,1,1
+6_M21A_M24A_010,6,M21,M24,1,0,10,10,2,8,8,2,0.8,5,1,1
+6_M21A_M23A_089,6,M21,M23,1,0,89,10,7,3,3,7,0.3,5,1,1
+6_M21A_M25A_053,6,M21,M25,1,0,53,10,6,4,4,6,0.4,5,1,1
+6_M21A_M25A_067,6,M21,M25,1,0,67,10,6,4,4,6,0.4,5,1,1
+6_M21A_M22A_036,6,M21,M22,1,0,36,10,5,5,5,5,0.5,5,1,1
+6_M21A_M22A_018,6,M21,M22,1,0,18,10,1,9,9,1,0.9,5,1,1
+6_M21A_M25A_043,6,M21,M25,1,0,43,10,3,7,7,3,0.7,5,1,1
+6_M21A_M23A_059,6,M21,M23,1,0,59,10,4,6,6,4,0.6,5,1,1
+6_M21A_M23A_038,6,M21,M23,1,0,38,10,1,9,9,1,0.9,5,1,1
+6_M21A_M25A_057,6,M21,M25,1,0,57,10,2,8,8,2,0.8,5,1,1
+6_M21A_M22A_028,6,M21,M22,1,0,28,10,1,9,9,1,0.9,5,1,1
+6_M21A_M23A_056,6,M21,M23,1,0,56,10,7,3,3,7,0.3,5,1,1
+6_M21A_M24A_079,6,M21,M24,1,0,79,10,4,6,6,4,0.6,5,1,1
+6_M21A_M23A_008,6,M21,M23,1,0,8,10,1,9,9,1,0.9,5,1,1
+6_M21A_M25A_069,6,M21,M25,1,0,69,10,8,2,2,8,0.2,5,1,1
+6_M21A_M22A_026,6,M21,M22,1,0,26,10,4,6,6,4,0.6,5,1,1
+6_M21A_M22A_076,6,M21,M22,1,0,76,10,6,4,4,6,0.4,5,1,1
+6_M21A_M25A_011,6,M21,M25,1,0,11,10,4,6,6,4,0.6,5,1,1
+6_M21A_M24A_030,6,M21,M24,1,0,30,10,4,6,6,4,0.6,5,1,1
+6_M21A_M23A_088,6,M21,M23,1,0,88,10,8,2,2,8,0.2,5,1,1
+6_M21A_M25A_045,6,M21,M25,1,0,45,10,7,3,3,7,0.3,5,1,1
+6_M21A_M23A_096,6,M21,M23,1,0,96,11,9,2,2,9,0.1818181818181818,5,1,1
+6_M21A_M22A_037,6,M21,M22,1,0,37,10,2,8,8,2,0.8,5,1,1
+6_M21A_M25A_054,6,M21,M25,1,0,54,10,7,3,3,7,0.3,5,1,1
+6_M21A_M23A_028,6,M21,M23,1,0,28,11,2,9,9,2,0.8181818181818182,5,1,1
+6_M21A_M24A_032,6,M21,M24,1,0,32,10,1,9,9,1,0.9,5,1,1
+6_M21A_M24A_084,6,M21,M24,1,0,84,10,5,5,5,5,0.5,5,1,1
+6_M21A_M23A_032,6,M21,M23,1,0,32,10,2,8,8,2,0.8,5,1,1
+6_M21A_M25A_055,6,M21,M25,1,0,55,10,5,5,5,5,0.5,5,1,1
+6_M21A_M25A_093,6,M21,M25,1,0,93,10,7,3,3,7,0.3,5,1,1
+6_M21A_M22A_024,6,M21,M22,1,0,24,10,3,7,7,3,0.7,5,1,1
+6_M21A_M24A_080,6,M21,M24,1,0,80,10,8,2,2,8,0.2,5,1,1
+6_M21A_M22A_056,6,M21,M22,1,0,56,10,2,8,8,2,0.8,5,1,1
+6_M21A_M23A_013,6,M21,M23,1,0,13,10,1,9,9,1,0.9,5,1,1
+6_M21A_M22A_007,6,M21,M22,1,0,7,10,2,8,8,2,0.8,5,1,1
+6_M21A_M24A_093,6,M21,M24,1,0,93,10,3,7,7,3,0.7,5,1,1
+6_M21B_M25B_067,6,M21,M25,1,0,67,10,6,4,4,6,0.4,5,1,1
+6_M21B_M22B_018,6,M21,M22,1,0,18,10,1,9,9,1,0.9,5,1,1
+6_M21B_M25B_093,6,M21,M25,1,0,93,11,4,7,7,4,0.6363636363636364,5,1,1
+6_M21B_M22B_024,6,M21,M22,1,0,24,10,1,9,9,1,0.9,5,1,1
+6_M21B_M25B_057,6,M21,M25,1,0,57,10,4,6,6,4,0.6,5,1,1
+6_M21B_M24B_061,6,M21,M24,1,0,61,10,5,5,5,5,0.5,5,1,1
+6_M21B_M24B_033,6,M21,M24,1,0,33,10,0,10,10,0,1.0,5,1,1
+6_M21B_M23B_089,6,M21,M23,1,0,89,10,9,1,1,9,0.1,5,1,1
+6_M21B_M23B_056,6,M21,M23,1,0,56,10,6,4,4,6,0.4,5,1,1
+6_M21B_M23B_028,6,M21,M23,1,0,28,10,3,7,7,3,0.7,5,1,1
+6_M21B_M25B_069,6,M21,M25,1,0,69,10,1,9,9,1,0.9,5,1,1
+6_M21B_M23B_008,6,M21,M23,1,0,8,10,2,8,8,2,0.8,5,1,1
+6_M21B_M23B_032,6,M21,M23,1,0,32,10,1,9,9,1,0.9,5,1,1
+6_M21B_M22B_077,6,M21,M22,1,0,77,10,4,6,6,4,0.6,5,1,1
+6_M21B_M25B_055,6,M21,M25,1,0,55,10,2,8,8,2,0.8,5,1,1
+6_M21B_M24B_079,6,M21,M24,1,0,79,10,3,7,7,3,0.7,5,1,1
+6_M21B_M25B_011,6,M21,M25,1,0,11,10,2,8,8,2,0.8,5,1,1
+6_M21B_M24B_080,6,M21,M24,1,0,80,10,7,3,3,7,0.3,5,1,1
+6_M21B_M22B_026,6,M21,M22,1,0,26,10,1,9,9,1,0.9,5,1,1
+6_M21B_M22B_056,6,M21,M22,1,0,56,11,1,10,10,1,0.9090909090909092,5,1,1
+6_M21B_M22B_007,6,M21,M22,1,0,7,10,1,9,9,1,0.9,5,1,1
+6_M21B_M25B_053,6,M21,M25,1,0,53,10,2,8,8,2,0.8,5,1,1
+6_M21B_M22B_037,6,M21,M22,1,0,37,10,3,7,7,3,0.7,5,1,1
+6_M21B_M25B_054,6,M21,M25,1,0,54,10,2,8,8,2,0.8,5,1,1
+6_M21B_M25B_043,6,M21,M25,1,0,43,10,5,5,5,5,0.5,5,1,1
+6_M21B_M24B_084,6,M21,M24,1,0,84,10,5,5,5,5,0.5,5,1,1
+6_M21B_M24B_074,6,M21,M24,1,0,74,10,5,5,5,5,0.5,5,1,1
+6_M21B_M22B_028,6,M21,M22,1,0,28,10,0,10,10,0,1.0,5,1,1
+6_M21B_M24B_093,6,M21,M24,1,0,93,10,6,4,4,6,0.4,5,1,1
+6_M21B_M25B_045,6,M21,M25,1,0,45,10,4,6,6,4,0.6,5,1,1
+6_M21B_M23B_006,6,M21,M23,1,0,6,10,3,7,7,3,0.7,5,1,1
+6_M21B_M23B_038,6,M21,M23,1,0,38,9,5,4,4,5,0.4444444444444444,5,1,1
+6_M21B_M22B_010,6,M21,M22,1,0,10,10,2,8,8,2,0.8,5,1,1
+6_M21B_M23B_096,6,M21,M23,1,0,96,10,3,7,7,3,0.7,5,1,1
+6_M21B_M23B_059,6,M21,M23,1,0,59,10,5,5,5,5,0.5,5,1,1
+6_M21B_M24B_032,6,M21,M24,1,0,32,10,5,5,5,5,0.5,5,1,1
+6_M21B_M22B_076,6,M21,M22,1,0,76,10,4,6,6,4,0.6,5,1,1
+6_M21B_M24B_030,6,M21,M24,1,0,30,10,2,8,8,2,0.8,5,1,1
+6_M21B_M23B_088,6,M21,M23,1,0,88,10,5,5,5,5,0.5,5,1,1
+6_M21B_M24B_010,6,M21,M24,1,0,10,10,3,7,7,3,0.7,5,1,1
+1_F16,1,F16,F16,0,1,100,37,36,1,36,1,0.972972972972973,4,2,1
+2_F16B,2,F16,F16,0,1,100,38,31,7,31,7,0.8157894736842105,4,2,1
+2_F16C,2,F16,F16,0,1,100,39,35,4,35,4,0.8974358974358975,4,2,1
+2_F16D,2,F16,F16,0,1,100,36,28,8,28,8,0.7777777777777778,4,2,1
+2_F16E,2,F16,F16,0,1,100,37,28,9,28,9,0.7567567567567568,4,2,1
+3_F16B,3,F16,F16,1,1,100,11,6,5,6,5,0.5454545454545454,4,2,1
+3_F16C,3,F16,F16,1,1,100,10,6,4,6,4,0.6,4,2,1
+3_F16D,3,F16,F16,1,1,100,10,6,4,6,4,0.6,4,2,1
+3_F16E,3,F16,F16,1,1,100,10,4,6,4,6,0.4,4,2,1
+4_F16_F17D,4,F16,F17,0,0,100,87,47,40,40,47,0.4597701149425287,4,2,1
+4_F16_F18E,4,F16,F18,0,0,100,33,18,15,15,18,0.4545454545454545,4,2,1
+4_F16_F19E,4,F16,F19,0,0,100,34,6,28,28,6,0.8235294117647058,4,2,1
+4_F16_F20C,4,F16,F20,0,0,100,38,1,37,37,1,0.9736842105263158,4,2,1
+5_F16_F17D,5,F16,F17,1,0,100,10,5,5,5,5,0.5,4,2,1
+5_F16_F18E,5,F16,F18,1,0,100,10,3,7,7,3,0.7,4,2,1
+5_F16_F19E,5,F16,F19,1,0,100,10,1,9,9,1,0.9,4,2,1
+5_F16_F20C,5,F16,F20,1,0,100,10,1,9,9,1,0.9,4,2,1
+6_F16A_F17A_071,6,F16,F17,1,0,71,10,7,3,3,7,0.3,4,2,1
+6_F16A_F18A_008,6,F16,F18,1,0,8,10,4,6,6,4,0.6,4,2,1
+6_F16A_F19A_048,6,F16,F19,1,0,48,10,5,5,5,5,0.5,4,2,1
+6_F16A_F19A_036,6,F16,F19,1,0,36,10,5,5,5,5,0.5,4,2,1
+6_F16A_F19A_070,6,F16,F19,1,0,70,10,8,2,2,8,0.2,4,2,1
+6_F16A_F18A_040,6,F16,F18,1,0,40,10,5,5,5,5,0.5,4,2,1
+6_F16A_F16A_100,6,F16,F16,1,0,100,10,5,5,5,5,0.5,4,2,1
+6_F16A_F20A_067,6,F16,F20,1,0,67,10,4,6,6,4,0.6,4,2,1
+6_F16A_F17A_016,6,F16,F17,1,0,16,10,6,4,4,6,0.4,4,2,1
+6_F16A_F20A_071,6,F16,F20,1,0,71,10,7,3,3,7,0.3,4,2,1
+6_F16A_F17A_009,6,F16,F17,1,0,9,10,5,5,5,5,0.5,4,2,1
+6_F16A_F20A_083,6,F16,F20,1,0,83,10,3,7,7,3,0.7,4,2,1
+6_F16A_F17A_081,6,F16,F17,1,0,81,10,8,2,2,8,0.2,4,2,1
+6_F16A_F17A_068,6,F16,F17,1,0,68,10,6,4,4,6,0.4,4,2,1
+6_F16A_F20A_027,6,F16,F20,1,0,27,10,2,8,8,2,0.8,4,2,1
+6_F16A_F19A_009,6,F16,F19,1,0,9,10,2,8,8,2,0.8,4,2,1
+6_F16A_F18A_003,6,F16,F18,1,0,3,10,6,4,4,6,0.4,4,2,1
+6_F16A_F17A_028,6,F16,F17,1,0,28,10,7,3,3,7,0.3,4,2,1
+6_F16A_F19A_030,6,F16,F19,1,0,30,10,2,8,8,2,0.8,4,2,1
+6_F16A_F17A_098,6,F16,F17,1,0,98,10,6,4,4,6,0.4,4,2,1
+6_F16A_F18A_064,6,F16,F18,1,0,64,10,6,4,4,6,0.4,4,2,1
+6_F16A_F20A_002,6,F16,F20,1,0,2,10,3,7,7,3,0.7,4,2,1
+6_F16A_F18A_053,6,F16,F18,1,0,53,10,9,1,1,9,0.1,4,2,1
+6_F16A_F19A_035,6,F16,F19,1,0,35,10,4,6,6,4,0.6,4,2,1
+6_F16A_F18A_066,6,F16,F18,1,0,66,10,6,4,4,6,0.4,4,2,1
+6_F16A_F19A_023,6,F16,F19,1,0,23,10,5,5,5,5,0.5,4,2,1
+6_F16A_F18A_088,6,F16,F18,1,0,88,10,7,3,3,7,0.3,4,2,1
+6_F16A_F17A_052,6,F16,F17,1,0,52,10,8,2,2,8,0.2,4,2,1
+6_F16A_F19A_089,6,F16,F19,1,0,89,10,6,4,4,6,0.4,4,2,1
+6_F16A_F18A_077,6,F16,F18,1,0,77,10,4,6,6,4,0.6,4,2,1
+6_F16A_F17A_007,6,F16,F17,1,0,7,10,9,1,1,9,0.1,4,2,1
+6_F16A_F20A_060,6,F16,F20,1,0,60,10,4,6,6,4,0.6,4,2,1
+6_F16A_F20A_066,6,F16,F20,1,0,66,10,7,3,3,7,0.3,4,2,1
+6_F16A_F19A_028,6,F16,F19,1,0,28,10,2,8,8,2,0.8,4,2,1
+6_F16A_F18A_075,6,F16,F18,1,0,75,10,2,8,8,2,0.8,4,2,1
+6_F16A_F20A_095,6,F16,F20,1,0,95,10,5,5,5,5,0.5,4,2,1
+6_F16A_F19A_060,6,F16,F19,1,0,60,10,3,7,7,3,0.7,4,2,1
+6_F16A_F20A_086,6,F16,F20,1,0,86,10,3,7,7,3,0.7,4,2,1
+6_F16A_F20A_084,6,F16,F20,1,0,84,10,4,6,6,4,0.6,4,2,1
+6_F16A_F18A_069,6,F16,F18,1,0,69,10,5,5,5,5,0.5,4,2,1
+6_F16A_F17A_086,6,F16,F17,1,0,86,10,7,3,3,7,0.3,4,2,1
+6_F16B_F19B_070,6,F16,F19,1,0,70,10,6,4,4,6,0.4,4,2,1
+6_F16B_F20B_060,6,F16,F20,1,0,60,10,1,9,9,1,0.9,4,2,1
+6_F16B_F19B_028,6,F16,F19,1,0,28,10,0,10,10,0,1.0,4,2,1
+6_F16B_F20B_095,6,F16,F20,1,0,95,10,5,5,5,5,0.5,4,2,1
+6_F16B_F17B_081,6,F16,F17,1,0,81,10,8,2,2,8,0.2,4,2,1
+6_F16B_F20B_083,6,F16,F20,1,0,83,10,4,6,6,4,0.6,4,2,1
+6_F16B_F18B_053,6,F16,F18,1,0,53,11,10,1,1,10,0.0909090909090909,4,2,1
+6_F16B_F17B_007,6,F16,F17,1,0,7,10,3,7,7,3,0.7,4,2,1
+6_F16B_F20B_002,6,F16,F20,1,0,2,10,3,7,7,3,0.7,4,2,1
+6_F16B_F19B_089,6,F16,F19,1,0,89,10,5,5,5,5,0.5,4,2,1
+6_F16B_F19B_009,6,F16,F19,1,0,9,10,3,7,7,3,0.7,4,2,1
+6_F16B_F17B_028,6,F16,F17,1,0,28,10,6,4,4,6,0.4,4,2,1
+6_F16B_F19B_048,6,F16,F19,1,0,48,10,3,7,7,3,0.7,4,2,1
+6_F16B_F18B_003,6,F16,F18,1,0,3,10,7,3,3,7,0.3,4,2,1
+6_F16B_F17B_052,6,F16,F17,1,0,52,10,5,5,5,5,0.5,4,2,1
+6_F16B_F17B_016,6,F16,F17,1,0,16,10,5,5,5,5,0.5,4,2,1
+6_F16B_F20B_027,6,F16,F20,1,0,27,10,3,7,7,3,0.7,4,2,1
+6_F16B_F18B_008,6,F16,F18,1,0,8,10,7,3,3,7,0.3,4,2,1
+6_F16B_F18B_040,6,F16,F18,1,0,40,10,7,3,3,7,0.3,4,2,1
+6_F16B_F20B_067,6,F16,F20,1,0,67,10,2,8,8,2,0.8,4,2,1
+6_F16B_F19B_023,6,F16,F19,1,0,23,10,3,7,7,3,0.7,4,2,1
+6_F16B_F18B_066,6,F16,F18,1,0,66,10,7,3,3,7,0.3,4,2,1
+6_F16B_F17B_071,6,F16,F17,1,0,71,10,5,5,5,5,0.5,4,2,1
+6_F16B_F19B_060,6,F16,F19,1,0,60,10,2,8,8,2,0.8,4,2,1
+6_F16B_F20B_086,6,F16,F20,1,0,86,10,4,6,6,4,0.6,4,2,1
+6_F16B_F18B_075,6,F16,F18,1,0,75,10,5,5,5,5,0.5,4,2,1
+6_F16B_F18B_088,6,F16,F18,1,0,88,10,7,3,3,7,0.3,4,2,1
+6_F16B_F20B_071,6,F16,F20,1,0,71,10,4,6,6,4,0.6,4,2,1
+6_F16B_F18B_077,6,F16,F18,1,0,77,10,4,6,6,4,0.6,4,2,1
+6_F16B_F18B_064,6,F16,F18,1,0,64,10,5,5,5,5,0.5,4,2,1
+6_F16B_F19B_036,6,F16,F19,1,0,36,10,5,5,5,5,0.5,4,2,1
+6_F16B_F17B_068,6,F16,F17,1,0,68,10,5,5,5,5,0.5,4,2,1
+6_F16B_F20B_066,6,F16,F20,1,0,66,11,4,7,7,4,0.6363636363636364,4,2,1
+6_F16B_F17B_086,6,F16,F17,1,0,86,10,9,1,1,9,0.1,4,2,1
+6_F16B_F18B_069,6,F16,F18,1,0,69,10,7,3,3,7,0.3,4,2,1
+6_F16B_F17B_098,6,F16,F17,1,0,98,10,6,4,4,6,0.4,4,2,1
+6_F16B_F19B_035,6,F16,F19,1,0,35,10,3,7,7,3,0.7,4,2,1
+6_F16B_F20B_084,6,F16,F20,1,0,84,10,6,4,4,6,0.4,4,2,1
+6_F16B_F17B_009,6,F16,F17,1,0,9,10,2,8,8,2,0.8,4,2,1
+6_F16B_F19B_030,6,F16,F19,1,0,30,10,3,7,7,3,0.7,4,2,1
+1_F06,1,F06,F06,0,1,100,29,26,3,26,3,0.896551724137931,2,2,1
+2_F06B,2,F06,F06,0,1,100,74,64,10,64,10,0.8648648648648649,2,2,1
+2_F06C,2,F06,F06,0,1,100,87,76,11,76,11,0.8735632183908046,2,2,1
+2_F06D,2,F06,F06,0,1,100,28,22,6,22,6,0.7857142857142857,2,2,1
+2_F06E,2,F06,F06,0,1,100,33,27,6,27,6,0.8181818181818182,2,2,1
+3_F06B,3,F06,F06,1,1,100,10,9,1,9,1,0.9,2,2,1
+3_F06C,3,F06,F06,1,1,100,10,5,5,5,5,0.5,2,2,1
+3_F06D,3,F06,F06,1,1,100,10,6,4,6,4,0.6,2,2,1
+3_F06E,3,F06,F06,1,1,100,10,7,3,7,3,0.7,2,2,1
+4_F06_F08C,4,F06,F08,0,0,100,32,15,17,17,15,0.53125,2,2,1
+4_F06_F09E,4,F06,F09,0,0,100,38,8,30,30,8,0.7894736842105263,2,2,1
+4_F06_F07E,4,F06,F07,0,0,100,74,20,54,54,20,0.7297297297297297,2,2,1
+4_F06_F10C,4,F06,F10,0,0,100,37,12,25,25,12,0.6756756756756757,2,2,1
+5_F06_F08C,5,F06,F08,1,0,100,10,2,8,8,2,0.8,2,2,1
+5_F06_F09E,5,F06,F09,1,0,100,10,1,9,9,1,0.9,2,2,1
+5_F06_F07E,5,F06,F07,1,0,100,10,2,8,8,2,0.8,2,2,1
+5_F06_F10C,5,F06,F10,1,0,100,10,2,8,8,2,0.8,2,2,1
+6_F06A_F08A_066,6,F06,F08,1,0,66,10,5,5,5,5,0.5,2,2,1
+6_F06A_F08A_075,6,F06,F08,1,0,75,10,6,4,4,6,0.4,2,2,1
+6_F06A_F09A_082,6,F06,F09,1,0,82,11,5,6,6,5,0.5454545454545454,2,2,1
+6_F06A_F06A_100,6,F06,F06,1,0,100,10,4,6,6,4,0.6,2,2,1
+6_F06A_F07A_056,6,F06,F07,1,0,56,10,3,7,7,3,0.7,2,2,1
+6_F06A_F07A_066,6,F06,F07,1,0,66,10,5,5,5,5,0.5,2,2,1
+6_F06A_F08A_011,6,F06,F08,1,0,11,10,1,9,9,1,0.9,2,2,1
+6_F06A_F07A_003,6,F06,F07,1,0,3,10,0,10,10,0,1.0,2,2,1
+6_F06A_F10A_014,6,F06,F10,1,0,14,10,2,8,8,2,0.8,2,2,1
+6_F06A_F09A_048,6,F06,F09,1,0,48,10,4,6,6,4,0.6,2,2,1
+6_F06A_F10A_093,6,F06,F10,1,0,93,10,6,4,4,6,0.4,2,2,1
+6_F06A_F10A_057,6,F06,F10,1,0,57,10,4,6,6,4,0.6,2,2,1
+6_F06A_F07A_014,6,F06,F07,1,0,14,10,0,10,10,0,1.0,2,2,1
+6_F06A_F08A_042,6,F06,F08,1,0,42,10,6,4,4,6,0.4,2,2,1
+6_F06A_F10A_050,6,F06,F10,1,0,50,10,0,10,10,0,1.0,2,2,1
+6_F06A_F07A_073,6,F06,F07,1,0,73,10,2,8,8,2,0.8,2,2,1
+6_F06A_F09A_065,6,F06,F09,1,0,65,10,6,4,4,6,0.4,2,2,1
+6_F06A_F07A_059,6,F06,F07,1,0,59,10,8,2,2,8,0.2,2,2,1
+6_F06A_F08A_031,6,F06,F08,1,0,31,10,6,4,4,6,0.4,2,2,1
+6_F06A_F10A_065,6,F06,F10,1,0,65,10,2,8,8,2,0.8,2,2,1
+6_F06A_F08A_096,6,F06,F08,1,0,96,10,6,4,4,6,0.4,2,2,1
+6_F06A_F08A_040,6,F06,F08,1,0,40,10,9,1,1,9,0.1,2,2,1
+6_F06A_F08A_018,6,F06,F08,1,0,18,10,5,5,5,5,0.5,2,2,1
+6_F06A_F10A_092,6,F06,F10,1,0,92,10,4,6,6,4,0.6,2,2,1
+6_F06A_F09A_067,6,F06,F09,1,0,67,10,8,2,2,8,0.2,2,2,1
+6_F06A_F09A_043,6,F06,F09,1,0,43,10,4,6,6,4,0.6,2,2,1
+6_F06A_F07A_027,6,F06,F07,1,0,27,10,6,4,4,6,0.4,2,2,1
+6_F06A_F09A_018,6,F06,F09,1,0,18,10,0,10,10,0,1.0,2,2,1
+6_F06A_F09A_047,6,F06,F09,1,0,47,10,5,5,5,5,0.5,2,2,1
+6_F06A_F10A_035,6,F06,F10,1,0,35,10,5,5,5,5,0.5,2,2,1
+6_F06A_F07A_035,6,F06,F07,1,0,35,10,3,7,7,3,0.7,2,2,1
+6_F06A_F10A_074,6,F06,F10,1,0,74,10,3,7,7,3,0.7,2,2,1
+6_F06A_F09A_010,6,F06,F09,1,0,10,10,3,7,7,3,0.7,2,2,1
+6_F06A_F10A_078,6,F06,F10,1,0,78,10,9,1,1,9,0.1,2,2,1
+6_F06A_F07A_057,6,F06,F07,1,0,57,10,6,4,4,6,0.4,2,2,1
+6_F06A_F09A_007,6,F06,F09,1,0,7,10,4,6,6,4,0.6,2,2,1
+6_F06A_F08A_082,6,F06,F08,1,0,82,10,4,6,6,4,0.6,2,2,1
+6_F06A_F07A_020,6,F06,F07,1,0,20,10,3,7,7,3,0.7,2,2,1
+6_F06A_F09A_098,6,F06,F09,1,0,98,10,7,3,3,7,0.3,2,2,1
+6_F06A_F10A_097,6,F06,F10,1,0,97,10,8,2,2,8,0.2,2,2,1
+6_F06A_F08A_050,6,F06,F08,1,0,50,10,6,4,4,6,0.4,2,2,1
+6_F06B_F09B_010,6,F06,F09,1,0,10,10,1,9,9,1,0.9,2,2,1
+6_F06B_F07B_073,6,F06,F07,1,0,73,11,1,10,10,1,0.9090909090909092,2,2,1
+6_F06B_F09B_018,6,F06,F09,1,0,18,10,1,9,9,1,0.9,2,2,1
+6_F06B_F07B_056,6,F06,F07,1,0,56,10,1,9,9,1,0.9,2,2,1
+6_F06B_F10B_014,6,F06,F10,1,0,14,11,3,8,8,3,0.7272727272727273,2,2,1
+6_F06B_F08B_066,6,F06,F08,1,0,66,11,9,2,2,9,0.1818181818181818,2,2,1
+6_F06B_F07B_035,6,F06,F07,1,0,35,10,2,8,8,2,0.8,2,2,1
+6_F06B_F08B_075,6,F06,F08,1,0,75,10,2,8,8,2,0.8,2,2,1
+6_F06B_F09B_067,6,F06,F09,1,0,67,10,2,8,8,2,0.8,2,2,1
+6_F06B_F07B_027,6,F06,F07,1,0,27,10,1,9,9,1,0.9,2,2,1
+6_F06B_F09B_082,6,F06,F09,1,0,82,10,3,7,7,3,0.7,2,2,1
+6_F06B_F08B_011,6,F06,F08,1,0,11,10,7,3,3,7,0.3,2,2,1
+6_F06B_F07B_003,6,F06,F07,1,0,3,10,3,7,7,3,0.7,2,2,1
+6_F06B_F10B_035,6,F06,F10,1,0,35,10,6,4,4,6,0.4,2,2,1
+6_F06B_F07B_020,6,F06,F07,1,0,20,10,1,9,9,1,0.9,2,2,1
+6_F06B_F10B_074,6,F06,F10,1,0,74,10,8,2,2,8,0.2,2,2,1
+6_F06B_F09B_065,6,F06,F09,1,0,65,10,2,8,8,2,0.8,2,2,1
+6_F06B_F09B_048,6,F06,F09,1,0,48,10,5,5,5,5,0.5,2,2,1
+6_F06B_F10B_078,6,F06,F10,1,0,78,10,1,9,9,1,0.9,2,2,1
+6_F06B_F10B_097,6,F06,F10,1,0,97,11,2,9,9,2,0.8181818181818182,2,2,1
+6_F06B_F09B_047,6,F06,F09,1,0,47,10,2,8,8,2,0.8,2,2,1
+6_F06B_F10B_092,6,F06,F10,1,0,92,10,2,8,8,2,0.8,2,2,1
+6_F06B_F08B_040,6,F06,F08,1,0,40,10,5,5,5,5,0.5,2,2,1
+6_F06B_F10B_065,6,F06,F10,1,0,65,10,1,9,9,1,0.9,2,2,1
+6_F06B_F09B_098,6,F06,F09,1,0,98,10,4,6,6,4,0.6,2,2,1
+6_F06B_F07B_066,6,F06,F07,1,0,66,10,0,10,10,0,1.0,2,2,1
+6_F06B_F09B_043,6,F06,F09,1,0,43,10,6,4,4,6,0.4,2,2,1
+6_F06B_F10B_050,6,F06,F10,1,0,50,10,0,10,10,0,1.0,2,2,1
+6_F06B_F07B_059,6,F06,F07,1,0,59,10,4,6,6,4,0.6,2,2,1
+6_F06B_F10B_093,6,F06,F10,1,0,93,10,3,7,7,3,0.7,2,2,1
+6_F06B_F08B_042,6,F06,F08,1,0,42,11,7,4,4,7,0.3636363636363636,2,2,1
+6_F06B_F07B_057,6,F06,F07,1,0,57,10,2,8,8,2,0.8,2,2,1
+6_F06B_F08B_018,6,F06,F08,1,0,18,10,4,6,6,4,0.6,2,2,1
+6_F06B_F08B_082,6,F06,F08,1,0,82,10,3,7,7,3,0.7,2,2,1
+6_F06B_F08B_031,6,F06,F08,1,0,31,10,7,3,3,7,0.3,2,2,1
+6_F06B_F08B_096,6,F06,F08,1,0,96,10,2,8,8,2,0.8,2,2,1
+6_F06B_F08B_050,6,F06,F08,1,0,50,10,3,7,7,3,0.7,2,2,1
+6_F06B_F07B_014,6,F06,F07,1,0,14,10,5,5,5,5,0.5,2,2,1
+6_F06B_F09B_007,6,F06,F09,1,0,7,10,0,10,10,0,1.0,2,2,1
+6_F06B_F10B_057,6,F06,F10,1,0,57,10,3,7,7,3,0.7,2,2,1