FlaxSpeaker

FlaxSpeaker (v0.5.0) is an end-to-end open-source speaker verification library built on JAX and Flax. Designed for education and academic research, it unifies dataset preparation, real-noise waveform augmentation (MUSAN + RIRS_NOISES + 3x speed perturbation) and zero-copy float16 memory-mapped feature shards, frame-level energy Voice Activity Detection (VAD) with peak Automatic Gain Control (AGC), an 80-D log-Mel frontend with CMVN, six neural sequence backbones (tiny, small, base, large), eight temporal pooling heads, ten metric-learning and angular-margin training objectives (including Sub-Center ArcFace K=3), three scoring paradigms (Cosine, PFAS, Dr-Vectors) with Top-300 AS-Norm and Quality-Aware Score Calibration (QMF), universal multi-domain stratified training, and one-command export to TFLite (FP32 / dynamic-range INT8), TensorFlow SavedModel, SafeTensors, and the Hugging Face Hub.

Two-Tier v0.5.0 Model Zoo (48 Pretrained Checkpoints = 44 Controlled Ablation Models + 4 Full-Data Universal Flagships).

1. Tier-2 Full-Data Universal Flagships (4 univ_full_* models, 6.98M–14.92M parameters): Trained on 12,193 physical speakers (36,579 speed-perturbed classes, 464,557 clean utterances / 1,857,149 augmented items across VoxCeleb2-dev + VoxCeleb1-dev + CN-Celeb1/2 + LibriSpeech-960h with MUSAN noise/music/babble and RIRS_NOISES room impulse responses) and evaluated on the full official trial lists (VoxCeleb1-O 37,611 trials, CN-Celeb-E 3,484,292 trials, and LibriSpeech test-clean 10,000 trials), reaching 13.92% VoxCeleb1-O QMF EER (univ_full_conformer_large_dr_vectors), 17.98% CN-Celeb-E Multi-Enroll EER (3-way fusion; 18.24% single-model), and 5.20% LibriSpeech test-clean 3-Utt EER (5.43% single-model).
2. Tier-1 Controlled Benchmark Zoo (44 univ_*, ls_*, cn_*, vox_* models, 0.63M–4.55M parameters): Trained to calibrated S_safe = X + 2 epoch step budgets (4,600 steps / 15.58 epochs on Universal 381 spk / 9,447 utts; 2,400–2,600 steps on LibriSpeech 232 spk; 1,600–1,800 steps on CN-Celeb2 119 spk; 1,000 steps on VoxCeleb1 30 spk) with uniform variable-length masked crops T ~ U(90, 180) (48 / 48 released checkpoints verified to 5e-5 step-0 loss tolerance as variable in pretrained_models/step0_fingerprints.json), Sub-Center ArcFace (K=3), and BatchNorm after activation in ECAPA-TDNN. The 44 Tier-1 models reach 13.94–28.91% pooled 1-utt EER (13.94–23.23% across the 19 univ_* models; 9.67–19.57% pooled 3-utt EER), with universal multi-domain training outperforming the matched LibriSpeech specialist in all 19 / 19 pairs. Multi-seed stability across seeds {0, 1, 2} has a pooled 1-utt EER standard deviation of 0.05% (univ_conformer_small_ge2e), 0.29% (univ_ecapa_tdnn_small_ge2e), and 0.41% (ls_conformer_small_ge2e; see RESEARCH_REPORT.md §10).

For the companion PyTorch educational project, see SpeakerRecognitionFromScratch.

---

Disclaimer & Open-Source Reproduction Notice

This is not an officially supported Google product.

This repository provides a from-scratch re-implementation in JAX / Flax, written using only the published papers (no proprietary code or data), of the mathematical formulations and neural architectures described in the following research papers:

1. Generalized End-to-End Loss for Speaker Verification — Li Wan, Quan Wang, Alan Papir, Ignacio Lopez Moreno (ICASSP 2018, arXiv:1710.10467)
2. Dr-Vectors: Decision Residual Networks and an Improved Loss for Speaker Recognition — Jason Pelecanos, Quan Wang, Ignacio Lopez Moreno (Interspeech 2021, arXiv:2104.01989)
3. Attentive Temporal Pooling for Conformer-based Streaming Language Identification in Long-form Speech — Quan Wang, Yang Yu, Jason Pelecanos, Yiling Huang, Ignacio Lopez Moreno (Odyssey 2022, arXiv:2202.12163) — a language identification paper whose attentive temporal pooling head is repurposed here for speaker embeddings
4. Parameter-Free Attentive Scoring for Speaker Verification — Jason Pelecanos, Quan Wang, Yiling Huang, Ignacio Lopez Moreno (Odyssey 2022, arXiv:2203.05642)

Because the original papers were trained and evaluated on proprietary internal datasets and infrastructure (for ~500k steps on thousands of speakers), exact numerical parity with the original paper tables is neither expected nor claimed. All pretrained checkpoints and benchmark tables in this repository are trained and evaluated from scratch on publicly available speech corpora (VoxCeleb2, VoxCeleb1, CN-Celeb1/2, and LibriSpeech) with speaker-disjoint splits.

---

System Architecture

1. Six Neural Sequence Backbones & Four Size Presets (0.63 M–14.92 M parameters across tiny, small, base, large)

All backbones share a unified (B, T, F) -> (B, T, D) sequence interface with exact variable-length padding mask invariance via ModernSpeakerEncoder (flaxspeaker/backbones.py) and scale automatically via size_variant: tiny | small | base | large:

• Conformer (conformer): Parallel Linear + 1D Conv Stem (Dense(D) + SiLU(LayerNorm(Conv1D(k=3)))) + Macaron FFN + Masked Multi-Head Self-Attention + Mask-Aware GroupNorm Depthwise Separable Convolution + Multi-Layer Feature Aggregation (MFA), with large preset (dim=384, num_layers=6, num_heads=6, ffn_dim=1536, subsample_factor=4, 6.98M–7.31M params).
• ECAPA-TDNN (ecapa_tdnn): 1D SE-Res2Net dilated multi-scale blocks with Multi-Layer Feature Aggregation (MFA), mask-aware BatchNorm running statistics after activation (use_batch_norm: true, norm_after_activation: true default in v0.5.0), and large preset (channels=1024, scale=8, mfa_dim=1536, se_bottleneck_dim=128, 14.92M params).
• Mamba (mamba): Selective State Space Model (SSM) implemented in pure JAX via parallel jax.lax.associative_scan (O(T) linear time complexity, supporting both causal unidirectional and mask-aware bidirectional bidirectional: true modes).
• ResNet (resnet): 2D Time-Frequency Residual Network with Squeeze-and-Excitation (SE-ResNet-18, 8 residual blocks across 4 stages (2, 2, 2, 2)).
• Transformer (transformer): Parallel Linear + 1D Conv Stem (Dense(D) + SiLU(LayerNorm(Conv1D(k=3)))) + Pre-LayerNorm Multi-Head Self-Attention encoder with Multi-Layer Feature Aggregation (MFA).
• LSTM (lstm): Deep stacked uni- or bidirectional LSTM (bidirectional: true, mask-aware with seq_lengths in both directions) with configurable hidden size, optional per-layer projection, and per-layer LayerNorm + residual connections.

2. Eight Temporal Pooling Heads (flaxspeaker/pooling.py)

• Attentive Temporal Pooling (atp): Sigmoid frame-gated statistics pooling (concat(mu_h, sigma_h)) extended from the single-head linear-sigmoid ATP of Wang et al., Odyssey 2022, arXiv:2202.12163 (a language-identification paper). FlaxSpeaker defaults add a tanh bottleneck (attention_dim: 128) and 4 channel-subspace heads (subspace_heads: true); attention_dim: 0, num_heads: 1 recovers the paper form. Followed by a nonlinear bottleneck FFN (LayerNorm -> Dense(256) -> LayerNorm -> ReLU -> Dense(d_e) -> L2_norm).
• Self-Attentive Pooling (sap) & Attentive Statistics Pooling (asp): Softmax attention-weighted frame mean and standard deviation.
• Mean Pooling (mean), Last-Frame Pooling (last), & Temporal Statistics Pooling (stats).
• Cumulative Statistics Pooling (cumulative_stats): Frame-wise causal running statistics for streaming inference.
• Parameter-Free Attentive Scoring Head (pfas): Supports attention_type: atp_ffn (the head of Pelecanos et al., Odyssey 2022, arXiv:2203.05642: ATP → affine + ReLU → linear → keys/values; univ_full_conformer_large_pfas uses the 16-config sweep winner pfas_paper_atpffn_k16_d32_v64_ln with K=16 tied keys × (d_k=32 + d_v=64) = 1,536-D, while odyssey2022_pfas_config(pure_paper_mode=True) uses 32 tied keys × (16 + 48) = 2,048-D), as well as attention_type: shared_nonlinear | unshared_nonlinear | dot | maxpool + use_global_anchor: true (PFAS-SN, used by the Tier-1 *_pfas* checkpoints: 4 keys × (32 + 128) = 640-D). All variants apply per-key L2 normalisation and global L2 normalisation of the concatenated value form (paper Eqs. 5–7).

3. Ten Speaker Recognition Losses & Three Scoring Backends (flaxspeaker/losses.py, flaxspeaker/scoring.py)

• Generalized End-to-End (GE2E) Loss: ge2e_softmax (Eq. 6, exact leave-one-out centroids) and ge2e_contrast (Eq. 7) with learnable scale w and bias b (Wan et al., ICASSP 2018). By default ge2e_contrast back-propagates Eq. 7 exactly; the released *_ge2e_contrast checkpoints use the hybrid contrast surrogate (ge2e_contrast_surrogate: true: forward value = Eq. 7, gradient = 0.25·Eq7 + 0.25·BCE + 0.50·GE2E-softmax-CE, ge2e_init_b: -1.0, ge2e_contrast_b_grad_scale: 0.5).
• Extended-Set Softmax Loss: extended_set_softmax, supporting both the paper's pure form (pure_extended_set: true, Eq. 2 of Pelecanos et al., Interspeech 2021) and the Blended Extended-Set Softmax (BXS, FlaxSpeaker variant) (pure_extended_set: false: 50/50 blend of row-softmax and extended-set softmax with a -log(N) logit shift).
• Decision Residual Networks (Dr-Vectors): scoring_type: dr_vector — single-stage end-to-end residual feed-forward scorer on [e_t, e_e, c]. interspeech2021_dr_vectors_config(pure_paper_mode=True) builds the paper's network (256-D embedding split into cosine_dims: 200 base + 56 side dims, 3 × Dense(256) leaky-ReLU(0.2) without normalisation, readout: weighted_sum, trainable_affine: true, pure extended-set softmax, reaching 18.30% pooled 1-utt EER at 4,600 steps in ablation_results.json). The released *_dr_vectors checkpoints use the compact down-scaled Dr-Vector variant (192-D or 256-D cosine, hidden (128, 64) ReLU + LayerNorm, linear read-out, residual_weight: 0.25, blended extended-set softmax).
• Parameter-Free Attentive Scoring (PFAS): scoring_type: pfas — cross-attention between test and enrollment (K, V) representations with per-trial softmax key–key attention weights A_{ij}, per-key and global L2 normalisation (Eqs. 5–7 of Pelecanos et al., Odyssey 2022), trainable scale, and key stacking (N = M × E) for multi-utterance enrollment.
• Angular Margin Classification Losses (with Sub-Center $K=3$ & Linear Margin Warm-Up):
  • Sub-Center ArcFace (subcenter_arcface / arcface with num_subcenters: 3): takes $\max_{k \in {1..K}} \cos(\theta_{i,j,k})$ per speaker class to absorb noisy labels before applying $\cos(\theta + m_t)$ with linear margin ramp m_t = m * min(1, t / margin_warmup_steps) (margin_warmup_steps = 500 on Universal, 300 on LibriSpeech, 250 on CN-Celeb2).
  • CosFace / AM-Softmax (cosface): cos(theta) - m_t with linear margin ramp margin_warmup_steps.
  • SphereFace / A-Softmax (sphereface): (-1)^k * cos(m * theta) - 2k with $\lambda$-annealing (1000 -> 5).
  • Standard Softmax (softmax), Pairwise Loss (pairwise), and Triplet Loss (triplet).
• Top-300 Adaptive Symmetric Score Normalization (AS-Norm) & Quality-Aware Score Calibration (QMF): Top-K (K=300) impostor-cohort z-normalisation (use_as_norm: true) with enforced trial-cohort disjointness (evaluated alongside raw scores), plus logistic Quality-Aware Score Calibration (qmf.py) combining raw/AS-Norm score, log duration frames, post-VAD speech ratio, and pre-L2 embedding norm.

4. End-to-End Audio Pipeline, Real Waveform Augmentation, float16 Shards & Model Export

• Frame-level energy VAD & AGC (flaxspeaker/vad.py): Frame energies are computed on the peak-normalised waveform (apply_agc = True, target peak 1.0) and the resulting mask is applied to the log-Mel frames (max(-42.0 dB, P_95(E) - 26.0 dB, P_10(E) + 4.5 dB), hangover smoothing min_speech_frames = 3, hangover_frames = 8, top-8-frame fallback).
• Acoustic Frontends, Full-Utterance Inference & float16 Memory-Mapped Shards (flaxspeaker/feature_extraction.py, flaxspeaker/sharded_cache.py): log_mel (natural-log Mel + energy VAD + CMVN), per-utterance variable sequence length sampling T ~ U(90, 180) frames across all 48 v0.5.0 checkpoints (variable_length_training: true) with exact binary padding masks (B, T), bucketed full-utterance inference (max_eval_seq_len = 0, capped at 3,000 frames / 30 s), and zero-copy np.memmap float16 feature shards (ShardedFeatureDataset).
• Real Waveform Augmentation & Feature-Domain Regularization (flaxspeaker/waveform_aug.py, flaxspeaker/specaug.py): Tier-2 full-data training applies 3x polyphase speed perturbation (0.9x, 1.0x, 1.1x -> 3x speaker classes) and online/offline MUSAN (speech, music, noise) + RIRS_NOISES room reverberation (MusanRirAugmenter), while Tier-1 controlled training applies mutually exclusive feature-domain noise (35%, SNR ~ U(12, 20) dB), SpecAugment (32.5%), and clean features (32.5%).
• Model Export (flaxspeaker/export.py, flaxspeaker/hf_compat.py): Export via jax2tf to TFLite (FP32, dynamic-range INT8) and TF SavedModel with a fixed, unmasked 180-frame (1.8 s) encoder window, and to SafeTensors / Hugging Face Hub (config.json, model.safetensors, flax_model.msgpack, eval_results.json, model card). Across all 16 exported .tflite files (8 models × FP32/INT8), dynamic-range INT8 quantization reduces file size by 3.44–3.88× (6.82–59.84 MB → 1.86–15.42 MB) while changing 180-frame verification EER by at most 0.00–0.14 percentage points relative to FP32 TFLite, with mean utterance-level cosine similarity cos(FP32, INT8) between 0.999972 and 0.999990 and a 1.85–1.93× CPU speedup on LSTM models (see RESEARCH_REPORT.md §11).

---

Installation

pip install -U flaxspeaker   # >= 0.5.0 recommended for the 48-model v0.5.0 zoo

To install from source: pip install git+https://github.com/wq2012/FlaxSpeaker.

For development, HuggingFace export, and TFLite conversion extras:

pip install -e ".[dev,export,hf]"

---

Quickstart & CLI Workflows

FlaxSpeaker provides standalone command-line scripts in scripts/ covering dataset preparation, shard generation, training, PFAS sweeps, evaluation, ablations, figure/table generation, and Hub deployment (in addition to python -m flaxspeaker):

1. Prepare Datasets & float16 Feature Shards (scripts/prepare_datasets.py / scripts/prepare_controlled_shards.py / scripts/prepare_full_corpora_shards.py)

# Prepare controlled academic datasets + combined Universal Multi-Domain split
python scripts/prepare_datasets.py --data_root ./data/speaker_datasets
python scripts/prepare_controlled_shards.py --data_root ./data/speaker_datasets

# Or download and shard the full 12,193-speaker multi-corpus datasets (VoxCeleb1/2 + CN-Celeb1/2 + LibriSpeech-960h + MUSAN + RIRS)
python scripts/download_full_corpora.py
python scripts/prepare_full_corpora_shards.py

2. Train Models (scripts/train.py / scripts/train_full_data_flagships.py)

Train from any converged YAML config (configs/*.yml) or run the Tier-2 full-data universal flagship pipeline:

# Train the top Tier-1 Universal Conformer + ATP + CosFace model (4,600 steps, margin_warmup_steps=500)
python scripts/train.py \
  --config configs/univ_conformer_small_cosface.yml \
  --train_csv ./data/speaker_datasets/splits/universal/train.csv \
  --eval_csv ./data/speaker_datasets/splits/universal/eval.csv \
  --trials_file ./data/speaker_datasets/splits/universal/trials.tsv \
  --output_dir ./checkpoints/univ_conformer_small_cosface

# Run the 16-config PFAS hyperparameter sweep and Tier-2 full-data universal flagships
python scripts/run_pfas_hparam_sweep.py
python scripts/train_full_data_flagships.py

3. Evaluate 1-Utt, 3-Utt, and VoxCeleb1-O-subset Trials (scripts/evaluate.py / scripts/eval_vox1_o_subset.py)

python scripts/evaluate.py \
  --model_dir wq2012/univ_full_conformer_large_arcface \
  --trials_file ./data/speaker_datasets/splits/universal/trials.tsv \
  --multi_trials_file ./data/speaker_datasets/splits/universal/multi_enroll_trials.tsv \
  --output_json eval_summary.json

# Evaluate all 48 models on the VoxCeleb1-O 10-speaker subset (writes pretrained_models/vox1_o_10spk_subset_results.json)
python scripts/eval_vox1_o_subset.py --data_root ./data/speaker_datasets --models_dir ./pretrained_models

4. Run End-to-End Speaker Verification Inference (scripts/inference.py)

python scripts/inference.py \
  --model_dir wq2012/univ_full_conformer_large_arcface \
  --enroll_audio hf_space/examples/librispeech_spk1_utt1.flac \
  --test_audio hf_space/examples/librispeech_spk1_utt2.flac \
  --apply_vad --vad_top_db 26.0 --full_sequence

5. Export to TFLite & Publish to the Hugging Face Hub (scripts/export_model.py / scripts/eval_tflite_models.py / scripts/update_hf_model_cards.py)

python scripts/export_model.py \
  --model_dir wq2012/univ_full_conformer_large_arcface \
  --output_dir ./exported_models/univ_full_conformer_large_arcface \
  --export_tflite \
  --quantize_int8

# Evaluate exported FP32 and INT8 TFLite models
python scripts/eval_tflite_models.py --data_root ./data/speaker_datasets --models_dir ./pretrained_models

# Regenerate all 48 model cards and push weights + TFLite + metadata to Hugging Face Hub
python scripts/update_hf_model_cards.py --push --include_weights

6. Reproduce Pilot Calibration, Full 48-Model Benchmark Suite, Ablations, Provenance & Paper Figures

# Phase-1 5,000-step convergence pilot & S_safe = X + 2 epoch schedule calibration
python scripts/run_phase1_convergence_pilot.py --data_root ./data/speaker_datasets

# Run the complete 44-model Tier-1 controlled benchmark matrix across all domains
python scripts/run_research_benchmarks.py --data_root ./data/speaker_datasets

# Reproduce controlled ablation studies + pure_paper_mode=True 4,600-step runs (ablation_results.json)
python scripts/run_r1_ablations.py --data_root ./data/speaker_datasets

# Multi-seed stability (3 seeds x 3 models), own-schedule horizons (400..4,600), matched-data controls
python scripts/run_r2_controls.py --data_root ./data/speaker_datasets

# Step-0 loss fingerprint of all 48 released checkpoints & training_provenance blocks
python scripts/fingerprint_released_checkpoints.py --data_root ./data/speaker_datasets
python scripts/write_training_provenance.py

# Regenerate Markdown / LaTeX benchmark tables and publication figures
python scripts/generate_report_tables.py --section all > /tmp/report_tables.md
python scripts/generate_paper_tables.py --paper_dir ../flax_speaker_paper
python scripts/generate_figures.py \
  --benchmark-json pretrained_models/benchmark_results.json \
  --output-dir ./figures

---

Python & HuggingFace API

Load any exported model directory or Hugging Face Hub repository ID (wq2012/<model_id>) with load_pretrained or FlaxSpeakerModel.from_pretrained:

import flaxspeaker
from flaxspeaker.configs import universal_multi_domain_config

# 1. Load a v0.5.0 full-data universal flagship or controlled checkpoint from Hugging Face Hub
verifier = flaxspeaker.load_pretrained(
    "wq2012/univ_full_conformer_large_arcface"
)

# 2. Verify two audio files directly (supports 1-utt or multi-utt enrollment)
score = verifier.verify(
    "hf_space/examples/librispeech_spk1_utt1.flac",
    "hf_space/examples/librispeech_spk1_utt2.flac",
)
print(f"Verification similarity score: {score:.4f}")

# 3. Or build a paper preset configuration programmatically
cfg = universal_multi_domain_config(
    backbone="conformer",
    size="large",
    loss_type="arcface",
    scoring_type="cosine",
)

---

Interactive Gradio Demo (hf_space/)

An interactive Gradio 5.50.0 web application is included in hf_space/. Hosting status: creating a Gradio Space under wq2012 on the Hugging Face Hub returns HTTP 402 (free CPU Spaces require a paid PRO subscription on that account), so run it locally (python hf_space/app.py) or deploy it to an account with Space creation enabled via python scripts/deploy_hf_space.py --repo_id <your-namespace>/FlaxSpeaker. Features:

• Real inference only (no precomputed outputs) across all 48 v0.5.0 models (4 Tier-2 univ_full_* Full-Data Universal Flagships + 19 Tier-1 univ_* + 25 Tier-1 Single-Domain ls_*, cn_*, vox_* checkpoints) from wq2012/FlaxSpeaker (Cosine, PFAS, and Dr-Vectors scoring).
• Automatic operating-point selection from each model's eval_results.json for 1-utterance vs. multi-utterance enrollment, configurable VAD thresholds (--vad_snr_margin_db, --vad_top_db), and steady-state JIT latency measurement.
• 1-utterance and 3-utterance enrollment with speech samples from LibriSpeech, CN-Celeb2, and VoxCeleb1, plus waveform + VAD gate, post-VAD log-Mel spectrogram, and embedding-profile visualizations.

python hf_space/app.py

---

48-Model Benchmark Summary & Pretrained Model Zoo (v0.5.0)

We trained and evaluated 48 v0.5.0 models across two tiers:

• Tier-2 Full-Data Universal Flagships (4 univ_full_* models): Trained on 12,193 speakers (36,579 speed-perturbed classes, 464,557 clean utterances / 1,857,149 augmented items across VoxCeleb2-dev + VoxCeleb1-dev + CN-Celeb1/2 + LibriSpeech-960h with MUSAN + RIRS_NOISES) and evaluated on the full official benchmarks (VoxCeleb1-O 37,611 trials, CN-Celeb-E 3,484,292 trials, LibriSpeech test-clean 10,000 trials) plus all controlled splits.
• Tier-1 Controlled Benchmark Zoo (44 models): 19 Universal Multi-Domain models (381 speakers / 9,447 utterances, 4,600 steps) + 25 Single-Domain specialist models (2,400–2,600 steps on LibriSpeech, 1,600–1,800 steps on CN-Celeb2, 1,000 steps on VoxCeleb1; total training compute 47.51 CPU hours, 5.8–397.3 min/model, median 36.2 min).

Key artifacts:

• Hugging Face Collection (48 Pretrained Models): wq2012/FlaxSpeaker
• Full technical report, convergence calibration, ablations & mathematical analysis: RESEARCH_REPORT.md
• Complete verified evaluation metrics JSON (48 models + 16 TFLite exports): pretrained_models/benchmark_results.json
• Tier-2 Full-Data Flagships & Official Benchmarks (VoxCeleb1-O, CN-Celeb-E, LibriSpeech test-clean + 2-way/3-way fusion): pretrained_models/flagship_official_results.json
• 16-Candidate Wide PFAS Hyperparameter Sweep JSON: pretrained_models/pfas_sweep_results.json
• Phase-1 5,000-step convergence pilot & S_safe schedule calibration JSON: pretrained_models/phase1_convergence_pilot.json
• Controlled ablation study & pure_paper_mode=True metrics JSON: pretrained_models/ablation_results.json
• VoxCeleb1-O 10-speaker subset evaluation JSON (48 models, 26 dB VAD): pretrained_models/vox1_o_10spk_subset_results.json
• Multi-seed stability (3 seeds × 3 models), own-schedule horizons (400..4,600), and matched-data controls (2,400 & 4,600 steps): pretrained_models/r2_seed_and_control_results.json
• Step-0 loss fingerprint of all 48 released checkpoints (48 / 48 variable): pretrained_models/step0_fingerprints.json
• Local Pretrained HuggingFace + SafeTensors + TFLite checkpoints: pretrained_models/

1. Tier-2 Full-Data Universal Flagships (12,193 Speakers) & Official Benchmarks (VoxCeleb1-O, CN-Celeb-E, LibriSpeech test-clean)

Hugging Face Repository	Backbone / Head / Objective	Params	Steps	VoxCeleb1-O Official (37,611 trials) Raw / AS-Norm / QMF EER (%)	CN-Celeb-E Official (3.48M trials) Single Raw / QMF / Multi-Enroll EER (%)	LibriSpeech test-clean (10,000 trials) 1-Utt Raw / QMF / 3-Utt EER (%)	Controlled 3-Domain Pooled 1-Utt / 3-Utt EER (%)	Controlled Macro-Avg 1-Utt / 3-Utt EER (%)
wq2012/univ_full_conformer_large_dr_vectors	conformer (large) / atp / extended_set_softmax (dr_vector)	7,148,743	2,400	14.32% / — / 13.92%	19.79% / 18.96% / 18.24%	9.13% / 8.52% / 5.60%	14.10% / 10.33% (raw 10.33%)	13.52% / 9.77%
wq2012/univ_full_conformer_large_pfas	conformer (large) / pfas / extended_set_softmax (pfas)	7,312,327	2,400	14.18% / — / 14.05%	19.78% / 18.83% / 18.25%	9.49% / 8.56% / 5.47%	14.19% / 10.20% (raw 10.20%)	13.36% / 9.43%
wq2012/univ_full_ecapa_c1024_arcface	ecapa_tdnn (large) / asp / arcface (cosine)	14,923,136	4,500	15.14% / 14.23% / 14.06%	21.18% / 19.67% / 18.89%	9.41% / 8.06% / 5.43%	14.69% / 11.00% (raw 11.03%)	14.30% / 10.67%
wq2012/univ_full_conformer_large_arcface	conformer (large) / atp / arcface (cosine)	6,983,364	4,500	15.20% / 14.56% / 14.58%	21.11% / 19.94% / 18.89%	9.90% / 8.73% / 6.13%	14.55% / 10.50% (raw 11.50%)	14.62% / 10.97%
2-Way Score Fusion (0.1·Conformer-L + 0.9·ECAPA-c1024)	Heterogeneous Cosine + AS-Norm + QMF	21,906,500	—	— / 14.18% / 13.98%	— / 19.59% / —	8.44% / 8.12% / —	—	—
3-Way Multi-Enroll Fusion (0.5·Conf-L + 0.3·ECAPA + 0.2·PFAS)	Cosine AS-Norm + Frame-Level PFAS	29,218,827	—	—	— / — / 17.98% (AUC 0.9098)	— / — / 5.20% (AUC 0.9871)	—	—

2. Original Published Paper Reference Results (Internal Proprietary Datasets)

Paper & Venue	Architecture & Loss / Scoring	Original Training Corpus	Original Evaluation Benchmark	Reported Metric in Original Paper
GE2E (Wan et al., ICASSP 2018)	3-layer LSTM (768 hidden, 256 proj) + GE2E-Softmax	~36 M utterances from ~18K speakers (anonymised logs)	Internal TI-SV: 1,000 speakers, ≈6.3 enrollment / 7.2 test utterances per speaker	3.55% EER (GE2E) vs. 4.06% (softmax) and 4.13% (TE2E), Table 3 — multi-utterance enrollment, not comparable to 1-utt numbers
Dr-Vectors (Pelecanos et al., Interspeech 2021)	LSTM + Extended-Set Softmax + Dr-Vectors Residual Scorer	Internal Proprietary Multi-Condition Corpus	Internal Multi-Condition Speaker Verification	~4–11% relative EER reduction over Cosine + GE2E
Conformer + ATP (Wang et al., Odyssey 2022)	Conformer Encoder + Attentive Temporal Pooling (ATP)	Internal Multi-Language Speech Corpus (65 languages)	Internal Streaming & Long-Form Identification	Significant EER & latency improvement over Mean/Last pooling
PFAS (Pelecanos et al., Odyssey 2022)	Conformer + Multi-Head (K, V) Parameter-Free Attentive Scoring	Internal Proprietary Multi-Vendor Corpus	Internal 1-Utt & Multi-Utt Speaker Verification	~8–14% relative EER reduction over Cosine baseline

3. Tier-1 Universal Multi-Domain Pretrained Models (19 univ_* Models, 4,600 Steps / 15.58 Epochs)

All 19 univ_* models below are trained simultaneously on LibriSpeech (232 spk) + CN-Celeb2 (119 spk) + VoxCeleb1 (30 spk) (381 speakers, 9,447 utterances) for 4,600 steps with Domain-Balanced Stratified Batching (75% within-domain, 25% cross-domain) and variable-length crops T ~ U(90, 180) (verified from pretrained_models/benchmark_results.json):

Hugging Face Repository	Steps	Backbone / Size	Pooling / Loss / Scoring	Params	LS 1-Utt / 3-Utt EER (%)	CN 1-Utt / 3-Utt EER (%)	VOX 1-Utt / 3-Utt EER (%)	Macro-Avg 1-Utt / 3-Utt EER (%)	Pooled 1-Utt / 3-Utt EER (%)	Macro 1-Utt / 3-Utt minDCF@0.01
wq2012/univ_conformer_small_cosface	4,600	conformer (small)	atp / cosface / cosine	2,029,892	7.67% / 3.80%	16.73% / 11.90%	17.05% / 11.80%	13.82% / 9.17%	13.94% / 9.67%	0.7159 / 0.5180
wq2012/univ_conformer_small_sphereface	4,600	conformer (small)	atp / sphereface / cosine	2,029,892	8.00% / 4.20%	17.00% / 12.00%	18.40% / 14.20%	14.47% / 10.13%	14.20% / 11.07%	0.7021 / 0.5100
wq2012/univ_conformer_small_arcface	4,600	conformer (small)	atp / arcface / cosine	2,029,892	8.27% / 4.60%	16.93% / 12.60%	18.05% / 12.30%	14.42% / 9.83%	14.80% / 10.10%	0.6943 / 0.6440
wq2012/univ_conformer_small_ge2e_contrast	4,600	conformer (small)	atp / ge2e_contrast / cosine	2,029,894	10.00% / 6.90%	18.63% / 14.10%	19.25% / 16.00%	15.96% / 12.33%	16.56% / 12.67%	0.7788 / 0.6440
wq2012/univ_resnet_small_ge2e	4,600	resnet (small)	asp / ge2e_softmax / cosine	2,239,922	11.03% / 5.70%	18.47% / 12.70%	21.45% / 17.90%	16.98% / 12.10%	16.79% / 12.57%	0.7926 / 0.8053
wq2012/univ_conformer_tiny_ge2e	4,600	conformer (tiny)	atp / ge2e_softmax / cosine	693,126	12.10% / 7.80%	18.73% / 15.00%	20.60% / 16.40%	17.14% / 13.07%	16.93% / 13.03%	0.9275 / 0.7800
wq2012/univ_ecapa_tdnn_base_ge2e	4,600	ecapa_tdnn (base)	asp / ge2e_softmax / cosine	3,190,626	11.73% / 7.90%	20.10% / 15.70%	20.20% / 17.20%	17.34% / 13.60%	17.25% / 14.00%	0.9028 / 0.7533
wq2012/univ_conformer_small_ge2e	4,600	conformer (small)	atp / ge2e_softmax / cosine	2,029,894	13.47% / 9.20%	19.40% / 16.30%	20.85% / 16.40%	17.91% / 13.97%	17.54% / 13.87%	0.9139 / 0.8227
wq2012/univ_conformer_small_pfas	4,600	conformer (small)	pfas / extended_set_softmax / pfas	2,162,187	11.73% / 8.10%	20.93% / 18.80%	19.45% / 16.20%	17.37% / 14.37%	17.60% / 14.70%	0.8726 / 0.7607
wq2012/univ_ecapa_tdnn_small_ge2e	4,600	ecapa_tdnn (small)	asp / ge2e_softmax / cosine	1,684,418	12.60% / 8.60%	19.93% / 15.40%	21.10% / 18.00%	17.88% / 14.00%	17.62% / 14.07%	0.9548 / 0.8473
wq2012/univ_ecapa_tdnn_tiny_ge2e	4,600	ecapa_tdnn (tiny)	asp / ge2e_softmax / cosine	626,722	13.00% / 8.50%	20.17% / 16.00%	20.10% / 16.30%	17.76% / 13.60%	17.80% / 14.20%	0.9285 / 0.8100
wq2012/univ_conformer_base_ge2e	4,600	conformer (base)	atp / ge2e_softmax / cosine	4,551,558	12.83% / 8.70%	19.53% / 15.60%	21.90% / 17.00%	18.09% / 13.77%	17.90% / 14.40%	0.9341 / 0.7893
wq2012/univ_conformer_small_pfas_ge2e	4,600	conformer (small)	pfas / ge2e_softmax / pfas	2,162,187	12.57% / 9.80%	20.60% / 17.90%	20.40% / 16.40%	17.86% / 14.70%	18.31% / 15.53%	0.8347 / 0.7593
wq2012/univ_conformer_small_ext_softmax	4,600	conformer (small)	atp / extended_set_softmax / cosine	2,029,894	13.03% / 9.00%	20.80% / 16.30%	21.50% / 16.40%	18.44% / 13.90%	18.61% / 14.93%	0.9206 / 0.7707
wq2012/univ_transformer_small_ge2e	4,600	transformer (small)	atp / ge2e_softmax / cosine	1,239,622	13.63% / 10.10%	21.43% / 16.50%	21.55% / 16.60%	18.87% / 14.40%	18.73% / 14.87%	0.8972 / 0.8613
wq2012/univ_lstm_small_ge2e	4,600	lstm (small)	asp / ge2e_softmax / cosine	1,710,914	14.17% / 9.60%	21.73% / 16.20%	21.10% / 17.30%	19.00% / 14.37%	19.14% / 14.87%	0.9560 / 0.9040
wq2012/univ_conformer_small_dr_vectors	4,600	conformer (small)	atp / extended_set_softmax / dr_vector	2,088,007	13.43% / 10.20%	21.93% / 16.40%	20.60% / 16.90%	18.66% / 14.50%	19.43% / 15.13%	0.9531 / 0.7827
wq2012/univ_mamba_small_ge2e	4,600	mamba (small)	asp / ge2e_softmax / cosine	1,028,674	16.23% / 12.20%	21.37% / 18.00%	20.95% / 17.40%	19.52% / 15.87%	19.57% / 16.57%	0.9644 / 0.9180
wq2012/univ_conformer_small_triplet	4,600	conformer (small)	atp / triplet / cosine	2,029,892	20.70% / 17.20%	24.87% / 20.00%	22.30% / 17.80%	22.62% / 18.33%	23.23% / 19.57%	0.9720 / 0.9213

4. Universal Multi-Domain vs. Single-Domain Training (19 / 19 Matched Pairs Improved)

At full convergence (S_safe), Universal Multi-Domain training outperforms the matched LibriSpeech specialist in all 19 / 19 univ_* vs. ls_* pairs across Pooled 1-Utt EER (-0.92 to -3.71 pp), Macro-Avg 1-Utt EER (-0.89 to -3.63 pp), and Pooled 3-Utt EER (-0.93 to -4.00 pp). Crucially, the matched-data control in r2_seed_and_control_results.json isolates speaker/domain diversity from utterance count: subsampling the 381-speaker Universal mix to 5,788 utterances (matching LibriSpeech) and training for 2,400 steps (matching compute) achieves 16.32% Pooled 1-Utt EER (16.64% Macro 1-Utt, 11.93% Pooled 3-Utt) vs. 20.20% for ls_conformer_small_ge2e (-3.88 pp improvement at identical utterance count and step budget), and 15.29% Pooled 1-Utt EER (11.20% Pooled 3-Utt) at 4,600 steps:

Model Paradigm	Representative Checkpoint (Steps)	LS 1-Utt / 3-Utt EER (%)	CN 1-Utt / 3-Utt EER (%)	VOX 1-Utt / 3-Utt EER (%)	Universal Macro-Avg 1-Utt / 3-Utt EER (%)	Universal Pooled 1-Utt / 3-Utt EER (%)
LibriSpeech-Only Specialist (GE2E)	wq2012/ls_conformer_small_ge2e (2,400)	12.07% / 7.50%	26.20% / 23.20%	21.25% / 18.50%	19.84% / 16.40%	20.20% / 16.93%
CN-Celeb2-Only Specialist (GE2E)	wq2012/cn_conformer_small_ge2e (1,600)	25.63% / 20.70%	21.93% / 17.50%	27.75% / 24.70%	25.11% / 20.97%	25.68% / 21.23%
VoxCeleb1-Only Specialist (GE2E)	wq2012/vox_conformer_small_ge2e (1,000)	20.00% / 15.20%	29.83% / 29.60%	20.50% / 16.00%	23.44% / 20.27%	23.64% / 20.70%
Universal Multi-Domain (GE2E)	wq2012/univ_conformer_small_ge2e (4,600)	13.47% / 9.20%	19.40% / 16.30%	20.85% / 16.40%	17.91% / 13.97%	17.54% / 13.87%
LibriSpeech-Only Specialist (ArcFace)	wq2012/ls_conformer_small_arcface (2,400)	8.13% / 4.20%	21.80% / 18.20%	17.65% / 13.60%	15.86% / 12.00%	16.23% / 12.40%
Universal Multi-Domain (ArcFace)	wq2012/univ_conformer_small_arcface (4,600)	8.27% / 4.60%	16.93% / 12.60%	18.05% / 12.30%	14.42% / 9.83%	14.80% / 10.10%
Universal Multi-Domain (CosFace)	wq2012/univ_conformer_small_cosface (4,600)	7.67% / 3.80%	16.73% / 11.90%	17.05% / 11.80%	13.82% / 9.17%	13.94% / 9.67%
Full-Data Universal Flagship (PFAS)	wq2012/univ_full_conformer_large_pfas (2,400)	8.07% / 4.20%	18.17% / 12.30%	15.85% / 11.80%	13.36% / 9.43%	14.19% / 10.20%

5. Tier-1 Single-Domain Pretrained Specialist Models (25 ls_*, cn_*, vox_* Models)

Hugging Face Repository	Train Domain (Steps)	Backbone / Size	Pooling / Loss / Scoring	Params	LS 1-Utt / 3-Utt EER (%)	CN 1-Utt / 3-Utt EER (%)	VOX 1-Utt / 3-Utt EER (%)	Macro-Avg 1-Utt / 3-Utt EER (%)	Pooled 1-Utt / 3-Utt EER (%)
wq2012/ls_conformer_small_cosface	librispeech (2,400)	conformer (small)	atp / cosface / cosine	2,029,892	8.03% / 4.10%	24.30% / 19.40%	19.00% / 13.60%	17.11% / 12.37%	17.62% / 13.67%
wq2012/ls_conformer_small_arcface	librispeech (2,400)	conformer (small)	atp / arcface / cosine	2,029,892	8.13% / 4.20%	21.80% / 18.20%	17.65% / 13.60%	15.86% / 12.00%	16.23% / 12.40%
wq2012/ls_conformer_small_sphereface	librispeech (2,400)	conformer (small)	atp / sphereface / cosine	2,029,892	8.40% / 4.10%	23.17% / 18.60%	16.40% / 12.80%	15.99% / 11.83%	16.60% / 12.13%
wq2012/ls_conformer_small_ge2e_contrast	librispeech (2,400)	conformer (small)	atp / ge2e_contrast / cosine	2,029,894	9.47% / 5.60%	26.27% / 22.00%	20.50% / 15.60%	18.74% / 14.40%	18.65% / 14.27%
wq2012/ls_resnet_small_ge2e	librispeech (2,600)	resnet (small)	asp / ge2e_softmax / cosine	2,239,922	9.53% / 6.00%	26.90% / 20.20%	22.60% / 19.10%	19.68% / 15.10%	20.12% / 15.87%
wq2012/ls_ecapa_tdnn_base_ge2e	librispeech (2,600)	ecapa_tdnn (base)	asp / ge2e_softmax / cosine	3,190,626	10.60% / 6.20%	26.87% / 22.10%	22.10% / 17.80%	19.86% / 15.37%	20.12% / 16.47%
wq2012/ls_conformer_tiny_ge2e	librispeech (2,400)	conformer (tiny)	atp / ge2e_softmax / cosine	693,126	11.07% / 6.30%	26.00% / 22.10%	20.90% / 17.70%	19.32% / 15.37%	19.76% / 15.73%
wq2012/ls_conformer_small_pfas	librispeech (2,400)	conformer (small)	pfas / extended_set_softmax / pfas	2,162,187	11.20% / 7.80%	26.87% / 23.80%	22.05% / 18.00%	20.04% / 16.53%	20.43% / 16.67%
wq2012/ls_ecapa_tdnn_small_ge2e	librispeech (2,400)	ecapa_tdnn (small)	asp / ge2e_softmax / cosine	1,684,418	11.37% / 6.10%	26.47% / 21.10%	21.10% / 18.10%	19.64% / 15.10%	20.05% / 15.77%
wq2012/ls_conformer_small_pfas_ge2e	librispeech (2,400)	conformer (small)	pfas / ge2e_softmax / pfas	2,162,187	11.60% / 9.20%	27.53% / 26.20%	20.90% / 17.00%	20.01% / 17.47%	20.57% / 17.70%
wq2012/ls_ecapa_tdnn_tiny_ge2e	librispeech (2,400)	ecapa_tdnn (tiny)	asp / ge2e_softmax / cosine	626,722	11.63% / 6.60%	27.80% / 22.50%	20.95% / 18.90%	20.13% / 16.00%	20.55% / 16.13%
wq2012/ls_conformer_small_ge2e	librispeech (2,400)	conformer (small)	atp / ge2e_softmax / cosine	2,029,894	12.07% / 7.50%	26.20% / 23.20%	21.25% / 18.50%	19.84% / 16.40%	20.20% / 16.93%
wq2012/ls_conformer_base_ge2e	librispeech (2,600)	conformer (base)	atp / ge2e_softmax / cosine	4,551,558	12.13% / 8.90%	27.30% / 22.50%	21.10% / 18.50%	20.18% / 16.63%	20.32% / 16.73%
wq2012/ls_conformer_small_dr_vectors	librispeech (2,400)	conformer (small)	atp / extended_set_softmax / dr_vector	2,088,007	13.20% / 8.90%	27.93% / 22.60%	21.80% / 17.90%	20.98% / 16.47%	21.40% / 16.93%
wq2012/ls_transformer_small_ge2e	librispeech (2,400)	transformer (small)	atp / ge2e_softmax / cosine	1,239,622	13.53% / 8.10%	28.20% / 24.20%	21.50% / 17.70%	21.08% / 16.67%	21.35% / 17.00%
wq2012/ls_lstm_small_ge2e	librispeech (2,400)	lstm (small)	asp / ge2e_softmax / cosine	1,710,914	14.00% / 10.20%	29.07% / 25.20%	22.20% / 19.00%	21.76% / 18.13%	21.73% / 18.27%
wq2012/ls_conformer_small_ext_softmax	librispeech (2,400)	conformer (small)	atp / extended_set_softmax / cosine	2,029,894	14.00% / 9.20%	29.40% / 25.80%	22.80% / 20.40%	22.07% / 18.47%	22.32% / 18.23%
wq2012/ls_mamba_small_ge2e	librispeech (2,400)	mamba (small)	asp / ge2e_softmax / cosine	1,028,674	14.43% / 10.40%	27.60% / 23.80%	21.15% / 18.20%	21.06% / 17.47%	21.38% / 17.63%
wq2012/ls_conformer_small_triplet	librispeech (2,400)	conformer (small)	atp / triplet / cosine	2,029,892	17.87% / 14.20%	32.10% / 28.70%	20.55% / 17.60%	23.51% / 20.17%	24.15% / 20.50%
wq2012/cn_ecapa_tdnn_small_arcface	cnceleb2 (1,800)	ecapa_tdnn (small)	asp / arcface / cosine	1,684,416	14.13% / 8.80%	17.63% / 13.00%	18.10% / 15.00%	16.62% / 12.27%	17.32% / 12.83%
wq2012/cn_conformer_small_ge2e	cnceleb2 (1,600)	conformer (small)	atp / ge2e_softmax / cosine	2,029,894	25.63% / 20.70%	21.93% / 17.50%	27.75% / 24.70%	25.11% / 20.97%	25.68% / 21.23%
wq2012/cn_conformer_small_pfas	cnceleb2 (1,600)	conformer (small)	pfas / extended_set_softmax / pfas	2,162,187	25.57% / 22.40%	22.20% / 20.70%	28.45% / 26.80%	25.41% / 23.30%	27.65% / 25.47%
wq2012/cn_conformer_small_dr_vectors	cnceleb2 (1,600)	conformer (small)	atp / extended_set_softmax / dr_vector	2,088,007	29.47% / 23.40%	24.53% / 20.80%	28.00% / 26.10%	27.33% / 23.43%	28.91% / 25.57%
wq2012/vox_conformer_small_ge2e	voxceleb1 (1,000)	conformer (small)	atp / ge2e_softmax / cosine	2,029,894	20.00% / 15.20%	29.83% / 29.60%	20.50% / 16.00%	23.44% / 20.27%	23.64% / 20.70%
wq2012/vox_ecapa_tdnn_small_ext_softmax	voxceleb1 (1,000)	ecapa_tdnn (small)	asp / extended_set_softmax / cosine	1,684,418	19.83% / 16.80%	30.97% / 30.00%	22.40% / 16.80%	24.40% / 21.20%	25.10% / 21.73%

---

Deviations from the Source Papers

Several released zoo configurations implement FlaxSpeaker variants of the cited algorithms alongside the exact paper-faithful configurations (icassp2018_ge2e_config, interspeech2021_dr_vectors_config, odyssey2022_conformer_atp_config, odyssey2022_pfas_config with pure_paper_mode=True). In v0.5.0, all four pure_paper_mode=True presets have been trained for 4,600 steps on the Universal corpus and benchmarked in pretrained_models/ablation_results.json (studies.paper_faithful_presets_smoke_test).

Component	Source paper	Released FlaxSpeaker v0.5.0 zoo checkpoints	Paper-faithful preset (4,600-step Universal Pooled 1 / 3-utt EER)
Training budget / data	GE2E: ~36 M utts, 18 K spk, N=64×M=10; Dr-Vectors/PFAS: 16 spk × 8 utts, 500 k steps	Tier-1 (44 models): 1,000–4,600 steps (S_safe) × (8 spk × 4 utts), 381 / 232 / 119 / 30 spk; Tier-2 (4 univ_full_* models): 2,400–4,500 steps on 12,193 speakers (464,557 clean utts / 1.86M augmented items); T ~ U(90, 180)	Evaluated on the 381-speaker Universal split (4,600 steps)
GE2E encoder	3 × LSTM-768 + 256-D projection, linear output	ls_lstm_* / univ_lstm_*: 3 × LSTM-256, no projection, per-layer LayerNorm + residual, ASP pooling, nonlinear bottleneck FFN (19.14% / 14.87%)	icassp2018_ge2e_config(pure_paper_mode=True) (30.52% / 29.13% — unnormalized 3×768 LSTM overfits/stalls on 9.4k utts)
GE2E contrast loss	Eq. 7	*_ge2e_contrast: hybrid contrast surrogate (value = Eq. 7; gradient = 0.25 Eq. 7 + 0.25 BCE + 0.50 softmax-CE), ge2e_init_b = −1, b grad scale 0.5 (16.56% / 12.67%)	ge2e_contrast_surrogate: false (default)
Extended-set softmax	Eq. 2 (pure)	*_ext_softmax, *_dr_vectors, *_pfas: blended 0.5 row-softmax + 0.5 extended-set with −log N shift (18.61% / 14.93%)	pure_extended_set: true (18.80% / 14.30% on Conformer-Small)
Dr-Vector scorer	200 base + 56 side dims; 3 × 256 leaky-ReLU(0.2); weighted sum; trainable affine	*_dr_vectors: cosine over all 192/256 dims, hidden (128, 64) ReLU + LayerNorm, linear read-out, fixed 0.25 residual weight (14.10% / 10.33% on univ_full_conformer_large_dr_vectors; 19.43% / 15.13% on Tier-1)	interspeech2021_dr_vectors_config(pure_paper_mode=True) (18.30% / 14.07%)
ATP pooling	single head, linear-sigmoid weights, full-D mean/std	attention_dim: 128 tanh bottleneck, 4 channel-subspace heads (subspace_heads: true); source paper is a language-ID paper	odyssey2022_conformer_atp_config(pure_paper_mode=True) (18.98% / 16.27%)
PFAS head	ATP → affine + ReLU(512) → linear; 32 tied keys (2,048-D)	univ_full_conformer_large_pfas: atp_ffn with K=16 tied keys × (32 + 64) = 1,536-D (14.19% / 10.20%); Tier-1 *_pfas*: PFAS-SN (shared_nonlinear + global ATP anchor, 4 keys × (32 + 128), 17.60% / 14.70%)	odyssey2022_pfas_config(pure_paper_mode=True) (23.02% / 21.00%)
ECAPA-TDNN / Angular Margin	BatchNorm / fixed margin or integer m = 4	BatchNorm after activation (17.62% / 14.07% vs 18.32% / 15.30% legacy LayerNorm before activation); Sub-Center ArcFace K=3 + linear margin_warmup_steps (13.94% CosFace, 14.80% ArcFace); real-valued SphereFace m = 1.35 with λ-annealing (14.20%)	—
Augmentation	waveform MTR with recorded noise + room simulator	Tier-2 univ_full_*: real MUSAN + RIRS_NOISES + 3x speed perturbation; Tier-1: synthetic log-Mel noise (SNR 12–20 dB) + SpecAugment	—
Evaluation	VoxCeleb1-O 37,611 trials (cleaned veri_test2.txt) over 40 speakers	Tier-2 univ_full_*: Official 37,611-trial VoxCeleb1-O (13.92% QMF), 3.48M-trial CN-Celeb-E (17.98% Multi-Enroll), 10,000-trial LibriSpeech test-clean (5.20% 3-Utt); Tier-1 + Tier-2: 10-speaker subset of veri_test2.txt (2,500 trials)	—

---

Testing, Linting & Documentation

Run the unit test suite (covering all backbones, exact padding-mask invariance across all backbone variants including BatchNorm ECAPA-TDNN, bidirectional LSTM, and Mamba, all losses including Sub-Center ArcFace, pure-vs-surrogate ge2e_contrast, pure/blended extended_set_softmax, and margin_warmup_steps, all pooling heads including "dot" and "unshared_nonlinear" PFAS attention, VAD + AGC, real waveform augmentation and float16 shards, Top-300 AS-Norm and QMF calibration, Dr-Vectors paper options, every pure_paper_mode preset field reaching the built module, strict config-key validation, export-window resolution, local-vs-Hub checkpoint equivalence, TFLite export, version consistency across pyproject.toml / setup.py / __init__.py / CHANGELOG.md, documentation script existence, and training_provenance consistency across all 48 checkpoints):

pytest -v tests.py
flake8 .
./run_pdoc.sh