FlaxSpeaker

FlaxSpeaker 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, 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, eight temporal pooling heads, nine metric-learning and angular-margin training objectives (seven of which are benchmarked), three scoring paradigms, universal multi-domain stratified training, and one-command export to TFLite (FP32 / dynamic-range INT8), TensorFlow SavedModel, SafeTensors, and the Hugging Face Hub.

Read this before using the released checkpoints. All 44 published models are 400-step, single-seed CPU demonstration checkpoints (32 utterances per step, ≈1.35 epochs over 9,447 utterances from 381 speakers). They reach 18.75–30.23 % pooled 1-utt EER on the project's trial lists and 22.5–32.2 % EER on a 10-speaker subset of VoxCeleb1-O (not comparable to published VoxCeleb1-O numbers; published systems reach ≈1 %). They are intended for relative comparisons of architectures and objectives under an identical budget, not as state-of-the-art models. Several released configurations are FlaxSpeaker variants of the cited algorithms (see Deviations from the source papers); the paper-faithful configurations are available as pure_paper_mode=True presets but have not been trained or released. Training-crop provenance. The released zoo, except the two *_ge2e_contrast checkpoints, was trained with fixed 180-frame crops and all-ones masks, not with the T ~ U(90, 180) variable-length crops that are the current pipeline default: the training code in use at the time defaulted the batchers to variable_length=False and the benchmark driver did not pass the flag. A step-0 loss fingerprint of all 44 checkpoints (scripts/fingerprint_released_checkpoints.py → pretrained_models/step0_fingerprints.json) reproduces the archived first-step loss to 4 decimals only under fixed crops for 34 checkpoints (plus 5 *_pfas* checkpoints within 7e-4, attributable to post-release PFAS scorer numerics), under variable-length crops for ls_conformer_small_ge2e_contrast, is not discriminable for 2 (univ_conformer_small_ge2e_contrast, univ_ecapa_tdnn_tiny_ge2e; assigned from git history), and cannot be established for the two *_triplet checkpoints (unseeded triplet sampler at the time). The shipped configs/*.yml and config.json now carry variable_length_training: false for the 40 fixed-crop checkpoints plus a training_provenance block. Evaluating the released weights with the current code reproduces the archived metrics exactly; re-training univ_conformer_small_ge2e (seed 42) with the current code and fixed crops gives 18.99 % [18.14, 19.82] pooled 1-utt EER vs. the archived 19.40 % [18.61, 20.26] (inside the released CI; ablation_results.json → variable_length_ablation.variable_length_false), whereas the variable-length default gives 20.94 % [20.02, 21.71] (r2_seed_and_control_results.json → seed42_rerun_current_code). See RESEARCH_REPORT.md §10 and pretrained_models/r3_fixed_crop_controls.json.

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, and the released 400-step checkpoints should not be read as reproductions of the papers' results. All pretrained checkpoints and benchmark tables in this repository are trained and evaluated from scratch on publicly available speech corpora (LibriSpeech, CN-Celeb2, and the 40-speaker VoxCeleb1 test partition) with speaker-disjoint splits.

---

System Architecture

1. Six Neural Sequence Backbones & Four Size Presets (0.63 M–4.55 M parameters in the released zoo; large is defined but never trained)

All backbones share a unified (B, T, F) -> (B, T, D) frame-rate-preserving 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).
• ECAPA-TDNN (ecapa_tdnn): 1D SE-Res2Net dilated multi-scale blocks with Multi-Layer Feature Aggregation (MFA).
• 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, 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. No dropout is implemented in any backbone.

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): Two representation heads. attention_type: atp_ffn is the head of Pelecanos et al., Odyssey 2022, arXiv:2203.05642 (ATP → affine + ReLU → linear → keys/values; odyssey2022_pfas_config(pure_paper_mode=True) uses 32 tied keys × (16 + 48) = 2,048-D). attention_type: shared_nonlinear + use_global_anchor: true (PFAS-SN, a FlaxSpeaker extension used by all released *_pfas* checkpoints) uses per-frame key/value projections with multi-head tanh frame attention anchored on a global ATP value summary, 4 keys × (32 + 128) = 640-D. Both apply per-key L2 normalisation and global L2 normalisation of the concatenated value form (paper Eqs. 5–7).

3. Nine 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 were trained with the hybrid contrast surrogate (ge2e_contrast_surrogate: true): the value is Eq. 7 but the gradient is 0.25·Eq7 + 0.25·BCE + 0.50·GE2E-softmax-CE, with ge2e_init_b: -1.0 and ge2e_contrast_b_grad_scale: 0.5 — a FlaxSpeaker objective, not Wan et al.'s contrast loss. When scoring_type is pfas or dr_vector, GE2E uses the even/odd split-batch test-vs-enrollment matrix instead of leave-one-out centroids.
• Extended-Set Softmax Loss: extended_set_softmax, supporting the paper's pure form (pure_extended_set: true, Eq. 2 of Pelecanos et al., Interspeech 2021) and the Blended Extended-Set Softmax (FlaxSpeaker variant) used by all released *_ext_softmax / *_dr_vectors / *_pfas checkpoints (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 (arXiv:2104.01989 Sec. 2/4: 256-D embedding split into cosine_dims: 200 base dims + 56 side-information dims, 3 × Dense(256) leaky-ReLU(0.2) without normalisation, readout: weighted_sum, score = cosine + network output, trainable_affine: true, trained with pure extended-set softmax). The released *_dr_vectors checkpoints are a down-scaled Dr-Vector variant (no base/side split, hidden (128, 64) ReLU + LayerNorm, linear read-out, fixed residual_weight: 0.25, blended extended-set softmax). The 3-term cosine expansion [c, c^2, sqrt(max(c, 0))] (num_cosine_terms: 3) is a FlaxSpeaker extension that does not appear in the paper (which has only two numbered equations).
• 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), a trainable scale, and key stacking (N = M × E) for multi-utterance enrollment.
• Angular Margin Classification Losses:
  • ArcFace / AAM-Softmax (arcface): cos(theta + m)
  • CosFace / AM-Softmax (cosface): cos(theta) - m
  • SphereFace / A-Softmax (sphereface): (-1)^k * cos(m * theta) - 2k
  • Standard Softmax (softmax), Pairwise Loss (pairwise), and Triplet Loss (triplet). (softmax and pairwise are implemented and unit-tested but no released checkpoint uses them.)
• Adaptive Symmetric Score Normalization (AS-Norm): Top-K impostor-cohort z-normalisation (use_as_norm: true). The cohort must be passed explicitly (cohort_embeddings=, e.g. training-speaker embeddings) and must be disjoint from the trial utterances; evaluate_verification_trials raises otherwise. No reported number uses AS-Norm.

4. End-to-End Audio Pipeline, Mutually Exclusive Augmentation, Variable-Length Training & Universal Multi-Domain Batching

• 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; 3-term dynamic thresholding max(-42.0 dB, P_95(E) - 26.0 dB, P_10(E) + 4.5 dB), optional spectral flatness <= 0.65 and speech-band (300-3400 Hz) energy ratio >= 0.25 gates, hangover smoothing (min_speech_frames = 3, hangover_frames = 8), and a top-8-frame fallback. One VAD constant (26 dB, the VadConfig() default) is used for data preparation, training, in-domain evaluation and the VoxCeleb1-O-subset evaluation.
• Acoustic Frontends & Variable-Length Training (flaxspeaker/feature_extraction.py): log_mel (used by every released checkpoint: natural-log Mel + energy VAD + CMVN; not reproducible with a Hugging Face AutoFeatureExtractor, which is why no preprocessor_config.json is shipped), optional whisper_mel / hf_compat / mfcc frontends, per-utterance variable sequence length sampling (T ~ U(min_seq_len, seq_len) = U(90, 180) frames, the pipeline default since commit 9accf95; the released zoo, except the two *_ge2e_contrast checkpoints, was trained with fixed 180-frame crops — see the provenance notice above and pretrained_models/step0_fingerprints.json) with exact binary padding masks (B, T), evaluation centre-cropping to max_eval_seq_len = 500 frames (5 s), and two-tier RAM + .npy disk caching (FeatureCache).
• Mutually Exclusive SpecAugment & feature-domain noise augmentation (flaxspeaker/specaug.py): The training-path "MTR" is synthetic log-Mel-domain augmentation — frequency-shaped Gaussian noise at SNR ~ U(12, 20) dB gated towards low-energy bins, plus a one-tap temporal smoothing with probability 0.4 — not waveform-domain multi-style training with recorded noise/RIR corpora. It is strictly mutually exclusive with SpecAugment (up to 2 frequency masks <= 8 mel bins and 2 time masks <= 12 frames): 35% noise only, 32.5% SpecAugment only, 32.5% clean. A standalone waveform helper (apply_waveform_augmentation) exists but is not on the training path.
• Universal Multi-Domain Stratified Batching (flaxspeaker/feature_extraction.py): Combines multiple heterogeneous datasets (LibriSpeech, CN-Celeb2, VoxCeleb1) using 75% within-domain hard-negative batches and 25% cross-domain score-calibration batches with temperature-balanced domain sampling p_d proportional to |S_d|^0.35.
• 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 input window (export.sequence_length; 0 = model.seq_len; the 12 released .tflite files use 180 frames = 1.8 s, encoder only, no scoring heads), and to SafeTensors / Hugging Face Hub directories (config.json, model.safetensors, eval_results.json, auto-generated model card). All 12 released .tflite files have been evaluated on the verification trial lists (scripts/eval_tflite_models.py): dynamic-range INT8 quantization reduces file size by 3.44–3.62× while changing 180-frame verification EER by at most 0.03–0.15 percentage points relative to FP32 TFLite (23.14% -> 23.08%, 25.39% -> 25.35%, 23.96% -> 23.81% on universal_pooled; 12.50% -> 12.37%, 16.30% -> 16.23%, 15.43% -> 15.40% on librispeech), with mean utterance-level cosine similarity cos(FP32, INT8) between 0.999919 and 0.999971 across all 1,988 unique evaluation utterances (see RESEARCH_REPORT.md §11).

---

Installation

pip install -U flaxspeaker   # >= 0.3.0 is required to load the released checkpoints

PyPI 0.2.0 was built on 2026-10-04 (git tag v0.2.0), before the 44 checkpoints and the load_pretrained / encode_features API existed; it cannot load any released checkpoint. 0.3.0 is the first release that can. To install the development head instead: 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 the full dataset preparation, training, evaluation, ablation, figure generation, and deployment lifecycle (in addition to python -m flaxspeaker):

1. Prepare Datasets & Multi-Domain Splits (scripts/prepare_datasets.py / scripts/prepare_data.py)

Generate speaker-disjoint train.csv, eval.csv, single-utterance trials.tsv, 3-utterance multi_enroll_trials.tsv, and precomputed feature caches for LibriSpeech, CN-Celeb2, VoxCeleb1, and the combined Universal multi-domain split:

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

# Or generate a synthetic smoke-test dataset split for local testing
python scripts/prepare_data.py --data_root ./data/synthetic --generate_synthetic

2. Train Models (scripts/train.py)

Train from a YAML config (configs/*.yml) or directly from one of the built-in paper preset configurations (universal_multi_domain, icassp2018_ge2e, interspeech2021_dr_vectors, odyssey2022_conformer_atp, odyssey2022_pfas):

# Train a Universal Multi-Domain Conformer + ATP + GE2E model
python scripts/train.py \
  --config configs/univ_conformer_small_ge2e.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_ge2e

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

Compute 1-utterance and 3-utterance EER (%), minDCF@0.01, minDCF@0.05, minDCF@0.10, ROC-AUC, and 95% bootstrap confidence intervals using any Hugging Face Hub model ID (wq2012/<model_id>) or local checkpoint directory, as well as the VoxCeleb1-O 10-speaker subset (official veri_test2.txt filtered to trials whose two speakers are both among the 10 held-out evaluation speakers, 2,500 trials, ≈72 % target; not comparable to published VoxCeleb1-O numbers — the other 30 of the 40 official test speakers are training data for vox_* and univ_* models):

python scripts/evaluate.py \
  --model_dir wq2012/univ_conformer_small_ge2e \
  --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 44 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)

Verify whether two (or more) audio files belong to the same speaker and optionally override VAD or full-sequence inference settings:

python scripts/inference.py \
  --model_dir wq2012/univ_conformer_small_ge2e \
  --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)

export_model.py writes a Hub-format directory (config.json, model.safetensors, flax_model.msgpack, model card, optional .tflite files) but does not push. Publishing is a separate step: scripts/update_hf_model_cards.py --push regenerates the 44 model cards and uploads README.md / config.json / eval_results.json to wq2012/<model_id>; weights are uploaded once with huggingface_hub.upload_folder (see that script's docstring). scripts/deploy_hf_space.py deploys the Gradio demo to a Space of your own.

python scripts/export_model.py \
  --model_dir wq2012/univ_conformer_small_ge2e \
  --output_dir ./exported_models/univ_conformer_small_ge2e \
  --export_tflite \
  --quantize_int8

# Evaluate exported FP32 and INT8 TFLite models on the 1-utterance trial lists
python scripts/eval_tflite_models.py --data_root ./data/speaker_datasets --models_dir ./pretrained_models

6. Reproduce Full 44-Model Benchmark Suite, Controlled Ablations, Provenance Checks & Paper Figures (scripts/run_research_benchmarks.py, scripts/run_r1_ablations.py, scripts/run_r2_controls.py, scripts/run_r3_fixed_crop_controls.py, scripts/fingerprint_released_checkpoints.py, scripts/generate_figures.py)

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

# Reproduce the 5 controlled ablation studies (pretrained_models/ablation_results.json)
python scripts/run_r1_ablations.py --data_root ./data/speaker_datasets

# Seed-noise floor (3 seeds x 2 models), honest 800-step horizon point, matched-data control
# (pretrained_models/r2_seed_and_control_results.json)
python scripts/run_r2_controls.py --data_root ./data/speaker_datasets

# Fixed-crop (T = 180) same-code controls: p_within in {0, 1}, 3-seed floor, ls seed-42 re-run
# (pretrained_models/r3_fixed_crop_controls.json)
python scripts/run_r3_fixed_crop_controls.py --data_root ./data/speaker_datasets

# Step-0 loss fingerprint of all 44 released checkpoints under fixed vs. variable-length crops
# (pretrained_models/step0_fingerprints.json), then write the training_provenance blocks
python scripts/fingerprint_released_checkpoints.py --data_root ./data/speaker_datasets
python scripts/write_training_provenance.py

# Regenerate the 2 paper figures (fig1_architecture.pdf and fig2_benchmarks.pdf) and README PNGs
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 pretrained universal multi-domain checkpoint from Hugging Face Hub or local path
verifier = flaxspeaker.load_pretrained(
    "wq2012/univ_conformer_small_ge2e"
)

# 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="small",
    loss_type="ge2e_softmax",
    scoring_type="cosine",
)

---

Interactive Gradio Demo (hf_space/)

An interactive Gradio 5.50.0 web application is included in hf_space/. It is not hosted on the Hugging Face Hub: creating a Gradio Space under wq2012 currently returns HTTP 402 (free CPU Spaces require a PRO subscription), so run it locally (python hf_space/app.py) or deploy it yourself with python scripts/deploy_hf_space.py --repo_id <your-namespace>/FlaxSpeaker. Features:

• Real inference only (no precomputed outputs) with a model selector across all 44 Universal Multi-Domain (univ_*) and Single-Domain (ls_*, cn_*, vox_*) checkpoints from wq2012/FlaxSpeaker (Cosine, PFAS, and Dr-Vectors scoring).
• The decision threshold defaults to the selected model's own pooled EER operating point from eval_results.json for the enrollment count actually used (1-utt for one enrollment utterance, 3-utt for two or three); untick the checkbox to use the manual slider instead (its range adapts to the scoring back-end). Both VAD knobs of the CLI are exposed (--vad_snr_margin_db, --vad_top_db). Latency is reported as steady-state model inference after JIT warm-up of the 1-, 2- and 3-utterance scoring kernels.
• 1-utterance and 3-utterance speaker enrollment with real speech samples from LibriSpeech, CN-Celeb2, and VoxCeleb1; waveform + VAD gate, post-VAD log-Mel spectrogram, and embedding-profile plots.
• AS-Norm is deliberately not offered (a web demo has no disjoint impostor cohort). hf_space/ is self-contained (requirements.txt installs flaxspeaker).

python hf_space/app.py

---

44-Model Benchmark Summary & Pretrained Model Zoo

We trained and evaluated 44 models (19 Universal Multi-Domain models trained jointly on LibriSpeech + CN-Celeb2 + VoxCeleb1 = 381 speakers / 9,447 utterances, plus 25 Single-Domain baseline models) across all three speaker-disjoint evaluation domains (8,000 1-utterance trials + 3,000 3-utterance trials + 2,500 VoxCeleb1-O-subset trials per model = 594,000 total trial evaluations). Every model: 400 steps × 32 utterances, seed 42, single run — see the seed-noise floor below before reading any sub-1 % difference as a ranking.

• Hugging Face Collection (44 Pretrained Models): wq2012/FlaxSpeaker
• Full technical report, ablations & mathematical analysis: RESEARCH_REPORT.md
• Complete verified evaluation metrics JSON (44 models + 12 TFLite exports): pretrained_models/benchmark_results.json
• Controlled ablation study metrics JSON (5 studies): pretrained_models/ablation_results.json
• VoxCeleb1-O 10-speaker subset evaluation JSON (44 models, 26 dB VAD matched to training): pretrained_models/vox1_o_10spk_subset_results.json
• Seed-noise floor over seeds {0, 1, 2}, 800-step horizon point with its own schedule, matched-data control (variable-length crops, current code): pretrained_models/r2_seed_and_control_results.json
• Fixed-crop (T = 180) same-code controls — p_within ∈ {0, 1}, 3-seed floor, ls_conformer_small_ge2e seed-42 re-run: pretrained_models/r3_fixed_crop_controls.json
• Step-0 loss fingerprint of all 44 released checkpoints (fixed vs. variable-length crops): pretrained_models/step0_fingerprints.json
• Local Pretrained HuggingFace + SafeTensors + TFLite checkpoints: pretrained_models/

1. 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 our 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	~11% relative EER reduction over Cosine + GE2E
Conformer + ATP (Wang et al., Odyssey 2022)	Conformer Encoder + Multi-Head 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

2. Universal Multi-Domain Pretrained Models (19 univ_* Models Across All 3 Domains)

All 19 univ_* models below are trained simultaneously on LibriSpeech (232 spk) + CN-Celeb2 (119 spk) + VoxCeleb1 (30 spk) (381 speakers, 9,447 utterances) with Domain-Balanced Stratified Batching (75% within-domain, 25% cross-domain) and evaluated on all three speaker-disjoint evaluation sets (verified from pretrained_models/benchmark_results.json):

Hugging Face Repository	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_tiny_ge2e	conformer (tiny)	atp / ge2e_softmax / cosine	693,126	13.23% / 8.80%	22.00% / 19.20%	22.55% / 17.60%	19.26% / 15.20%	18.75% / 15.03%	0.9211 / 0.8020
wq2012/univ_conformer_small_ge2e	conformer (small)	atp / ge2e_softmax / cosine	2,029,894	12.50% / 8.80%	22.33% / 20.00%	24.80% / 19.70%	19.88% / 16.17%	19.40% / 16.03%	0.9359 / 0.8433
wq2012/univ_conformer_small_dr_vectors	conformer (small)	atp / extended_set_softmax / dr_vector	2,088,007	15.20% / 9.90%	22.33% / 18.20%	22.75% / 17.50%	20.09% / 15.20%	20.10% / 15.40%	0.9348 / 0.7500
wq2012/univ_conformer_small_ext_softmax	conformer (small)	atp / extended_set_softmax / cosine	2,029,894	14.83% / 9.80%	22.40% / 18.80%	23.30% / 19.90%	20.18% / 16.17%	20.50% / 16.17%	0.9635 / 0.8973
wq2012/univ_conformer_base_ge2e	conformer (base)	atp / ge2e_softmax / cosine	4,551,558	15.37% / 11.10%	22.40% / 18.50%	22.85% / 19.90%	20.21% / 16.50%	19.85% / 16.80%	0.9763 / 0.8347
wq2012/univ_conformer_small_pfas_ge2e	conformer (small)	pfas / ge2e_softmax / pfas	2,162,187	15.20% / 11.00%	21.93% / 17.40%	23.65% / 19.80%	20.26% / 16.07%	19.86% / 16.27%	0.9741 / 0.8727
wq2012/univ_mamba_small_ge2e	mamba (small)	asp / ge2e_softmax / cosine	1,028,674	14.20% / 9.50%	23.20% / 18.50%	23.75% / 17.70%	20.38% / 15.23%	20.41% / 16.07%	0.9571 / 0.7920
wq2012/univ_transformer_small_ge2e	transformer (small)	atp / ge2e_softmax / cosine	1,239,622	15.17% / 10.70%	22.67% / 19.00%	23.95% / 18.00%	20.59% / 15.90%	20.00% / 16.33%	0.9694 / 0.8427
wq2012/univ_conformer_small_pfas	conformer (small)	pfas / extended_set_softmax / pfas	2,162,187	15.83% / 10.80%	23.17% / 18.60%	22.95% / 18.00%	20.65% / 15.80%	20.35% / 16.00%	0.8967 / 0.8060
wq2012/univ_conformer_small_ge2e_contrast	conformer (small)	atp / ge2e_contrast / cosine	2,029,894	14.73% / 8.30%	23.27% / 17.10%	24.45% / 19.40%	20.82% / 14.93%	20.69% / 15.33%	0.9396 / 0.8613
wq2012/univ_lstm_small_ge2e	lstm (small)	asp / ge2e_softmax / cosine	1,710,914	14.97% / 11.20%	25.53% / 21.60%	22.50% / 19.50%	21.00% / 17.43%	21.31% / 18.00%	0.9620 / 0.9173
wq2012/univ_ecapa_tdnn_base_ge2e	ecapa_tdnn (base)	asp / ge2e_softmax / cosine	3,190,626	18.63% / 12.20%	25.97% / 21.20%	22.20% / 17.30%	22.27% / 16.90%	22.81% / 18.20%	0.9760 / 0.9340
wq2012/univ_ecapa_tdnn_small_ge2e	ecapa_tdnn (small)	asp / ge2e_softmax / cosine	1,684,418	19.00% / 11.80%	25.13% / 22.40%	23.30% / 18.80%	22.48% / 17.67%	23.16% / 18.00%	0.9889 / 0.9120
wq2012/univ_ecapa_tdnn_tiny_ge2e	ecapa_tdnn (tiny)	asp / ge2e_softmax / cosine	626,722	19.07% / 12.50%	25.27% / 21.40%	23.35% / 17.20%	22.56% / 17.03%	23.02% / 18.00%	0.9613 / 0.8933
wq2012/univ_conformer_small_triplet	conformer (small)	atp / triplet / cosine	2,029,892	18.50% / 14.40%	26.67% / 24.00%	25.40% / 22.00%	23.52% / 20.13%	24.07% / 20.33%	0.9477 / 0.9247
wq2012/univ_conformer_small_arcface	conformer (small)	atp / arcface / cosine	2,029,892	20.57% / 17.40%	25.27% / 21.40%	25.00% / 21.80%	23.61% / 20.20%	25.11% / 21.67%	0.9869 / 0.9160
wq2012/univ_conformer_small_cosface	conformer (small)	atp / cosface / cosine	2,029,892	20.37% / 17.10%	25.73% / 22.20%	26.40% / 20.30%	24.17% / 19.87%	25.60% / 21.83%	0.9914 / 0.9193
wq2012/univ_conformer_small_sphereface	conformer (small)	atp / sphereface / cosine	2,029,892	22.27% / 18.10%	25.67% / 23.40%	25.15% / 20.40%	24.36% / 20.63%	25.84% / 22.07%	0.9858 / 0.9347
wq2012/univ_resnet_small_ge2e	resnet (small)	asp / ge2e_softmax / cosine	2,239,922	24.53% / 21.40%	29.33% / 26.40%	28.25% / 25.80%	27.37% / 24.53%	27.89% / 25.47%	0.9987 / 0.9847

3. Universal Multi-Domain vs. Single-Domain Training (matched Conformer-Small / GE2E)

Universal training lowers pooled 1-utt EER relative to the matched LibriSpeech specialist in 11 of the 19 univ_* vs. ls_* pairs (every Conformer metric-learning objective incl. the contrast surrogate, Dr-Vector, PFAS-SN and BXS variants, plus Transformer, Mamba and ECAPA-TDNN-base with GE2E) and raises it in the other 8 (LSTM, ECAPA-TDNN-small/tiny, ResNet-18, and the Conformer ArcFace/CosFace/SphereFace/triplet models); it is worse on macro 1-utt EER in 8 of 19 pairs (Transformer, ECAPA-TDNN-small/tiny, ResNet-18, ArcFace, CosFace, SphereFace, triplet). It never wins in-domain (the LibriSpeech specialist is 9.67 % vs. 12.50 % on LibriSpeech; the VoxCeleb1 specialist 20.80 % vs. 24.80 % on VoxCeleb1). The comparison also confounds domain diversity with 1.6× more training data (9,447 vs. 5,788 utterances); see the matched-data control in r2_seed_and_control_results.json. All of these are single-seed differences; see the seed-noise floor in RESEARCH_REPORT.md §10. Note that the released checkpoints were trained with fixed 180-frame crops (above), and that on identical code at 400 steps fixed crops score about 2 pp better pooled 1-utt EER than the variable-length default (18.99 % vs. 20.94 %, seed 42; 3-seed means 19.58 ± 0.76 % vs. 20.64 ± 0.73 %; r3_fixed_crop_controls.json, r2_seed_and_control_results.json).

Model Paradigm	Representative Checkpoint	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	wq2012/ls_conformer_small_ge2e	9.67% / 5.50%	28.50% / 26.10%	22.00% / 17.50%	20.06% / 16.37%	20.12% / 16.73%
CN-Celeb2-Only Specialist	wq2012/cn_conformer_small_ge2e	20.80% / 16.00%	21.33% / 16.80%	27.20% / 23.90%	23.11% / 18.90%	23.93% / 20.13%
VoxCeleb1-Only Specialist	wq2012/vox_conformer_small_ge2e	16.33% / 11.70%	27.87% / 25.80%	20.80% / 16.40%	21.67% / 17.97%	22.40% / 17.63%
Universal Multi-Domain (small)	wq2012/univ_conformer_small_ge2e	12.50% / 8.80%	22.33% / 20.00%	24.80% / 19.70%	19.88% / 16.17%	19.40% / 16.03%
Universal Multi-Domain (tiny)	wq2012/univ_conformer_tiny_ge2e	13.23% / 8.80%	22.00% / 19.20%	22.55% / 17.60%	19.26% / 15.20%	18.75% / 15.03%

4. Single-Domain Pretrained Baseline Models (25 ls_*, cn_*, vox_* Models)

Hugging Face Repository	Train Domain	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_ge2e	librispeech	conformer (small)	atp / ge2e_softmax / cosine	2,029,894	9.67% / 5.50%	28.50% / 26.10%	22.00% / 17.50%	20.06% / 16.37%	20.12% / 16.73%
wq2012/ls_conformer_small_ext_softmax	librispeech	conformer (small)	atp / extended_set_softmax / cosine	2,029,894	10.03% / 6.20%	29.87% / 25.50%	24.10% / 20.20%	21.33% / 17.30%	21.05% / 17.53%
wq2012/ls_conformer_small_pfas	librispeech	conformer (small)	pfas / extended_set_softmax / pfas	2,162,187	10.50% / 6.20%	28.83% / 24.90%	22.90% / 18.70%	20.74% / 16.60%	20.88% / 16.80%
wq2012/ls_conformer_small_pfas_ge2e	librispeech	conformer (small)	pfas / ge2e_softmax / pfas	2,162,187	10.50% / 6.20%	28.87% / 25.00%	23.40% / 20.20%	20.92% / 17.13%	20.69% / 17.17%
wq2012/ls_conformer_tiny_ge2e	librispeech	conformer (tiny)	atp / ge2e_softmax / cosine	693,126	10.53% / 6.20%	30.00% / 25.80%	25.70% / 21.60%	22.08% / 17.87%	21.91% / 18.07%
wq2012/ls_transformer_small_ge2e	librispeech	transformer (small)	atp / ge2e_softmax / cosine	1,239,622	10.60% / 6.20%	28.10% / 23.30%	22.30% / 16.90%	20.33% / 15.47%	20.80% / 16.37%
wq2012/ls_conformer_base_ge2e	librispeech	conformer (base)	atp / ge2e_softmax / cosine	4,551,558	10.77% / 6.60%	28.87% / 23.90%	24.65% / 19.90%	21.43% / 16.80%	21.26% / 17.00%
wq2012/ls_conformer_small_dr_vectors	librispeech	conformer (small)	atp / extended_set_softmax / dr_vector	2,088,007	11.17% / 6.20%	28.80% / 24.00%	22.30% / 19.60%	20.76% / 16.60%	20.69% / 16.60%
wq2012/ls_conformer_small_ge2e_contrast	librispeech	conformer (small)	atp / ge2e_contrast / cosine	2,029,894	11.43% / 7.10%	29.57% / 25.10%	25.75% / 20.40%	22.25% / 17.53%	22.80% / 19.07%
wq2012/ls_mamba_small_ge2e	librispeech	mamba (small)	asp / ge2e_softmax / cosine	1,028,674	12.17% / 7.90%	29.13% / 23.30%	21.05% / 16.10%	20.78% / 15.77%	21.30% / 17.10%
wq2012/ls_ecapa_tdnn_tiny_ge2e	librispeech	ecapa_tdnn (tiny)	asp / ge2e_softmax / cosine	626,722	12.60% / 8.40%	30.87% / 26.40%	21.70% / 17.30%	21.72% / 17.37%	22.10% / 17.67%
wq2012/ls_lstm_small_ge2e	librispeech	lstm (small)	asp / ge2e_softmax / cosine	1,710,914	12.77% / 7.90%	28.03% / 24.90%	23.35% / 19.70%	21.38% / 17.50%	21.25% / 17.73%
wq2012/ls_ecapa_tdnn_small_ge2e	librispeech	ecapa_tdnn (small)	asp / ge2e_softmax / cosine	1,684,418	13.83% / 9.10%	31.70% / 28.40%	19.65% / 15.90%	21.73% / 17.80%	22.48% / 17.83%
wq2012/ls_conformer_small_cosface	librispeech	conformer (small)	atp / cosface / cosine	2,029,892	14.13% / 10.80%	30.53% / 25.00%	23.00% / 17.40%	22.56% / 17.73%	23.26% / 18.17%
wq2012/ls_conformer_small_arcface	librispeech	conformer (small)	atp / arcface / cosine	2,029,892	14.47% / 9.60%	28.73% / 25.50%	23.85% / 18.30%	22.35% / 17.80%	22.80% / 18.23%
wq2012/ls_ecapa_tdnn_base_ge2e	librispeech	ecapa_tdnn (base)	asp / ge2e_softmax / cosine	3,190,626	14.47% / 10.60%	33.33% / 30.20%	19.90% / 17.00%	22.57% / 19.27%	23.29% / 18.90%
wq2012/ls_conformer_small_sphereface	librispeech	conformer (small)	atp / sphereface / cosine	2,029,892	14.73% / 11.00%	29.93% / 23.10%	22.20% / 16.30%	22.29% / 16.80%	23.29% / 17.97%
wq2012/ls_resnet_small_ge2e	librispeech	resnet (small)	asp / ge2e_softmax / cosine	2,239,922	15.20% / 11.30%	32.17% / 29.70%	21.40% / 20.20%	22.92% / 20.40%	23.94% / 20.47%
wq2012/ls_conformer_small_triplet	librispeech	conformer (small)	atp / triplet / cosine	2,029,892	16.07% / 13.50%	28.93% / 25.20%	22.00% / 20.20%	22.33% / 19.63%	23.41% / 19.80%
wq2012/cn_conformer_small_dr_vectors	cnceleb2	conformer (small)	atp / extended_set_softmax / dr_vector	2,088,007	23.00% / 16.20%	21.17% / 16.70%	26.30% / 20.40%	23.49% / 17.77%	23.95% / 18.97%
wq2012/cn_conformer_small_ge2e	cnceleb2	conformer (small)	atp / ge2e_softmax / cosine	2,029,894	20.80% / 16.00%	21.33% / 16.80%	27.20% / 23.90%	23.11% / 18.90%	23.93% / 20.13%
wq2012/cn_conformer_small_pfas	cnceleb2	conformer (small)	pfas / extended_set_softmax / pfas	2,162,187	24.47% / 18.90%	22.13% / 19.80%	28.15% / 24.80%	24.92% / 21.17%	26.20% / 23.20%
wq2012/cn_ecapa_tdnn_small_arcface	cnceleb2	ecapa_tdnn (small)	asp / arcface / cosine	1,684,416	30.53% / 23.40%	28.07% / 23.80%	31.40% / 31.20%	30.00% / 26.13%	30.23% / 26.47%
wq2012/vox_conformer_small_ge2e	voxceleb1	conformer (small)	atp / ge2e_softmax / cosine	2,029,894	16.33% / 11.70%	27.87% / 25.80%	20.80% / 16.40%	21.67% / 17.97%	22.40% / 17.63%
wq2012/vox_ecapa_tdnn_small_ext_softmax	voxceleb1	ecapa_tdnn (small)	asp / extended_set_softmax / cosine	1,684,418	20.07% / 14.60%	32.40% / 30.60%	21.10% / 17.80%	24.52% / 21.00%	24.57% / 20.97%

---

Deviations from the Source Papers

The released checkpoints implement FlaxSpeaker variants of the cited algorithms. The paper-faithful configurations are available as presets (icassp2018_ge2e_config, interspeech2021_dr_vectors_config, odyssey2022_conformer_atp_config, odyssey2022_pfas_config with pure_paper_mode=True; every field they set is consumed by the built module and covered by tests.py), but no paper-faithful preset has been trained or released, so the table below describes what was actually run.

Component	Source paper	Released FlaxSpeaker checkpoints	Paper-faithful preset
Training budget / data	GE2E: ~36 M utts, 18 K spk, N=64×M=10; Dr-Vectors/PFAS: 16 spk × 8 utts, 500 k steps	400 steps × (8 spk × 4 utts); ≤ 25 utts/spk; 381 (universal) / 232 / 119 / 30 speakers; single seed	not available (single CPU workstation)
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	icassp2018_ge2e_config(pure_paper_mode=True)
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 gradient scale 0.5	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	pure_extended_set: true
Dr-Vector scorer	200 base + 56 side dims; 3 × 256 leaky-ReLU(0.2); weighted sum; trainable affine	*_dr_vectors: cosine over all 192 dims, hidden (128, 64) ReLU + LayerNorm, linear read-out, fixed 0.25 residual weight. The 3-term [c, c², √max(c,0)] expansion is a FlaxSpeaker extension (ablation only)	interspeech2021_dr_vectors_config(pure_paper_mode=True)
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	attention_dim: 0, num_heads: 1
PFAS head	ATP → affine + ReLU(512) → linear; 32 tied keys (2,048-D)	*_pfas*: PFAS-SN — per-frame key/value projections, multi-head tanh frame attention + global ATP anchor, 4 keys × (32 + 128)	odyssey2022_pfas_config(pure_paper_mode=True) (32 × (16 + 48); dims inferred from the paper's totals)
Conformer	12 × 512, 4–8 heads, kernel 32, relative PE, dropout, stack-by-2	2–4 blocks × 128–256, 4–8 heads, kernel 15, no PE, GroupNorm(2) instead of BatchNorm, conv stem + MFA, no dropout	odyssey2022_conformer_atp_config(pure_paper_mode=True) (conv-before-MHSA, sinusoidal PE, stack-by-2 after block 3)
ECAPA-TDNN / SphereFace	BatchNorm / integer m = 4	LayerNorm / real-valued m = 1.35 with λ-annealing	—
Augmentation	waveform MTR with recorded noise + room simulator	synthetic log-Mel noise (SNR 12–20 dB) + one-tap smoothing	—
Evaluation	VoxCeleb1-O 37,720 trials over 40 speakers after VoxCeleb1-dev / VoxCeleb2 training	10-speaker subset of veri_test2.txt (2,500 trials, 72 % target); 30/40 official test speakers used for training	—

---

Testing, Linting & Documentation

Run the unit test suite (covering all backbones, exact padding-mask invariance on all seven backbone variants including bidirectional LSTM, losses including pure-vs-surrogate ge2e_contrast gradients and pure/blended extended_set_softmax, pooling heads, VAD + AGC, mutually exclusive augmentation, Dr-Vectors paper options, every pure_paper_mode preset field reaching the built module, strict config-key validation, AS-Norm cohort disjointness, export-window resolution, HF compatibility, TFLite export, version consistency across pyproject.toml / setup.py / __init__.py, that every scripts/*.py named in the documentation exists, and that all 44 config.json files carry a training_provenance block consistent with step0_fingerprints.json):

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