FlaxSpeaker: univ_conformer_small_arcface (conformer - small)

Disclaimer & Reproduction Notice

This is not an officially supported Google product.

This checkpoint and the accompanying FlaxSpeaker library are a from-scratch re-implementation in JAX/Flax, written using only the published papers (no proprietary code or data), of speaker recognition algorithms from:

• GE2E: Generalized End-to-End Loss for Speaker Verification (arXiv:1710.10467)
• Dr-Vectors: Decision Residual Networks and an Improved Loss for Speaker Recognition (arXiv:2104.01989)
• ATP (pooling head repurposed from a language identification paper): Attentive Temporal Pooling for Conformer-based Streaming Language Identification in Long-form Speech (arXiv:2202.12163)
• PFAS: Parameter-Free Attentive Scoring for Speaker Verification (arXiv:2203.05642)

The experiments in those papers were run on Google-internal infrastructure and data, not with this library. This is a converged small-corpus benchmark checkpoint (4,600 steps / ~31.16 epochs on 9,447 utterances from 381 speakers), not a full-corpus reproduction of the papers' large-scale results -- see Training budget and Deviations from the source papers below.

Training Budget (read this first)

This model was trained to convergence for 4,600 gradient steps (S_safe = X + 2 epochs safety margin, warmup 460 steps) with mini-batches of 64 utterances (8 speakers x 4 utterances), i.e. ~31.16 epochs over the 9,447-utterance 3-domain controlled training pool (381 speakers), on a CPU workstation with seed 42. Angular margin warmup is linearly ramped over the first 500 steps (margin_warmup_steps) before holding the target margin. Across the training trajectory, Universal Pooled EER progressed from step 400: 28.12% (1-utt) / 24.93% (3-utt) -> step 4000 (knee): 14.59% (1-utt) / 10.00% (3-utt) -> step 4600 (final): 14.80% (1-utt) / 10.10% (3-utt). While training loss and validation EER have reached their small-corpus convergence plateau on this 381-speaker open split (<= 25 utterances/speaker), absolute EERs remain bounded by training speaker diversity compared to industrial benchmarks trained on thousands of speakers (see below).

Training-crop provenance. This checkpoint was trained with U(90, 180)-frame variable-length crops and an exact validity mask (the current pipeline default); the shipped config.json / YAML (variable_length_training: true) describes the released weights. Step-0 loss fingerprint at the released seed (scripts/fingerprint_released_checkpoints.py, pretrained_models/step0_fingerprints.json): archived 9.3033; current code with fixed 180-frame crops 9.1251, with U(90, 180) crops 9.3033 (match: variable). Evaluating the released weights with the current code reproduces the archived metrics in eval_results.json. At the converged S_safe budget, the seed-to-seed standard deviation of pooled 1-utt EER across seeds 0-2 is 0.08-0.41 % absolute (pretrained_models/r2_seed_and_control_results.json), while the 8,000-trial 95 % bootstrap CI half-width is ~0.8 %.

For orientation only: published ECAPA-TDNN systems trained on VoxCeleb2-dev reach about 0.9% EER on the full VoxCeleb1-O list (Desplanques et al., 2020), and sub-1M-parameter systems about 2% (Wang et al., EfficientTDNN, 2022). The checkpoints released here are trained to convergence (including a +2-epoch safety margin) on the self-contained 381-speaker (<=25 utterances/speaker) benchmark corpus rather than VoxCeleb2-dev (5,994 speakers, ~1.1M utterances), and are intended for controlled relative comparisons within the FlaxSpeaker benchmark, not as state-of-the-art models trained on thousand-speaker corpora.

Model Summary

Property	Value
Model ID	univ_conformer_small_arcface
Hugging Face Repository	wq2012/univ_conformer_small_arcface
Hugging Face Collection	FlaxSpeaker Collection
GitHub Repository	wq2012/FlaxSpeaker
PyPI Package	flaxspeaker
Backbone Architecture	conformer (small)
Temporal Pooling	atp
Training Objective	arcface (arcface)
Verification Scorer	Cosine similarity (cosine)
Output Representation Dim	192
Inference Parameters (encoder + scoring head; excludes training-only classifier weights)	2,029,892
Training Data	LibriSpeech + CN-Celeb2 + VoxCeleb1 (universal multi-domain, 381 speakers, 9,447 utterances)
Training Steps / Batch	4,600 steps x 64 utterances (~31.16 epochs), seed 42
Audio Frontend	80-dim natural-log Mel @ 16000 Hz + frame-level energy VAD + CMVN (flaxspeaker.AudioFrontend; not a Hugging Face AutoFeatureExtractor)

Model Files Included

• model.safetensors: model weights (encoder + GE2E w,b / Dr-Vector / PFAS heads) in SafeTensors format
• flax_model.msgpack: native Flax Msgpack serialized parameters
• config.json: FlaxSpeakerConfig (full ExperimentConfig, incl. the frontend/VAD settings)
• eval_results.json: multi-domain evaluation metrics (1-utt & 3-utt EER, minDCF, ROC-AUC, 95% bootstrap CIs, VoxCeleb1-O 10-speaker subset)

Quickstart Usage

Python API (from Hugging Face Hub or Local Directory)

Requires flaxspeaker >= 0.5.0 (pip install -U flaxspeaker).

from flaxspeaker import load_pretrained

model = load_pretrained("wq2012/univ_conformer_small_arcface")
emb = model.extract_embedding("path/to/utterance.flac")
score = model.verify("path/to/enroll.flac", "path/to/test.flac")
print(f"Embedding shape: {emb.shape}, Verification score: {score:.4f}")

Command-Line Inference

python scripts/inference.py --model_dir wq2012/univ_conformer_small_arcface \
    --enroll_audio path/to/enroll.flac --test_audio path/to/test.flac

Multi-Domain Evaluation Results

1-Utterance Verification Across Evaluation Domains

Trial lists are 50/50 target/impostor over held-out speakers (LibriSpeech 40, CN-Celeb2 30, VoxCeleb1 10 speakers); minDCF@0.01 is close to its trivial ceiling of 1.0 for every model under this trial mix.

Evaluation Split	Trials	Raw EER (%)	AS-Norm EER (%)	95% Bootstrap CI (%)	minDCF@0.01	minDCF@0.10	ROC-AUC
LibriSpeech (English Read)	3,000	8.27%	6.97%	[7.23, 9.32]	0.4647	0.3467	0.9727
CN-Celeb2 (Chinese Multi-Genre)	3,000	16.93%	16.20%	[15.42, 18.20]	0.8593	0.6600	0.9150
VoxCeleb1 (In-the-Wild Celebrity)	2,000	18.05%	17.40%	[16.40, 19.93]	0.7590	0.7030	0.9149
Universal Pooled (All 3 Domains)	8,000	14.80%	13.28%	[13.93, 15.47]	0.7135	0.5580	0.9381
Universal Macro-Average (3 Domains)	—	14.42%	13.52%	—	0.6943	0.5699	0.9342
VoxCeleb1-O 10-speaker subset (see caveat)	2,500	19.53%	17.97%	[17.83, 21.14]	0.7390	0.7361	0.9010

Caveat: VoxCeleb1-O 10-speaker subset: the official veri_test2.txt list filtered to trials whose enrollment and test 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 used as training data by vox_* and univ_* models.

3-Utterance Multi-Enrollment Verification Across Evaluation Domains

Evaluation Split	Trials	3-Utt EER (%)	95% Bootstrap CI (%)	minDCF@0.01	ROC-AUC
LibriSpeech (3-Utt)	1,000	4.60%	[3.20, 5.80]	0.2760	0.9905
CN-Celeb2 (3-Utt)	1,000	12.60%	[10.20, 14.20]	0.7860	0.9537
VoxCeleb1 (3-Utt)	1,000	12.30%	[10.45, 14.75]	0.8700	0.9457
Universal Pooled (3-Utt)	3,000	10.10%	[8.96, 11.23]	0.7387	0.9661
Universal Macro-Average (3-Utt)	—	9.83%	—	0.6440	0.9633

Single seed; the 95% bootstrap CIs resample trials (not speakers). Differences below about 1% absolute EER between FlaxSpeaker checkpoints are within the seed-to-seed variation measured in pretrained_models/r2_seed_and_control_results.json (see the GitHub README) and should not be interpreted as rankings.

Deviations from the Source Papers

• Training regime: 4,600 gradient steps x 64 utterances (N = 8 speakers x M = 4, ~31.16 epochs) on <= 25 utterances/speaker from 381 speakers (9,447 utterances); the source papers train for ~500k steps with N = 16-64 / M = 8-10 on thousands of speakers.
• Augmentation: corpus-backed MUSAN environmental noise, music, and multi-speaker babble + RIRS_NOISES room impulse responses + 3-way polyphase speed perturbation (0.9x, 1.0x, 1.1x) + SpecAugment.
• Conformer: parallel linear + Conv1D stem, no positional encoding, 3 blocks x 192-D, 4 heads, kernel 15, mask-aware GroupNorm(2), no dropout, multi-layer feature aggregation; the Google papers use 12 x 512 blocks, 4-8 heads, kernel 32, relative positional encoding and dropout.
• ATP pooling: 4 heads with a tanh bottleneck (128-D) and subspace_heads = True; the papers use a single head with a linear-sigmoid weight (attention_dim = 0). Note that arXiv:2202.12163 is a language-identification paper whose pooling head is repurposed here.
• Frontend: 80-D natural-log Mel + frame-level energy VAD (26 dB dynamic range) + CMVN; evaluation processes up to 3000 frames (30.0 s).

Citation

@inproceedings{wan2018generalized,
  title={Generalized End-to-End Loss for Speaker Verification},
  author={Wan, Li and Wang, Quan and Papir, Alan and Moreno, Ignacio Lopez},
  booktitle={IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  pages={4879--4883},
  year={2018}
}

@inproceedings{pelecanos2021dr,
  title={Dr-Vectors: Decision Residual Networks and an Improved Loss for Speaker Recognition},
  author={Pelecanos, Jason and Wang, Quan and Moreno, Ignacio Lopez},
  booktitle={Interspeech},
  pages={4603--4607},
  year={2021}
}

@inproceedings{wang2022attentive,
  title={Attentive Temporal Pooling for Conformer-based Streaming Language Identification in Long-form Speech},
  author={Wang, Quan and Yu, Yang and Pelecanos, Jason and Huang, Yiling and Moreno, Ignacio Lopez},
  booktitle={The Speaker and Language Recognition Workshop (Odyssey)},
  year={2022}
}

@inproceedings{pelecanos2022parameter,
  title={Parameter-Free Attentive Scoring for Speaker Verification},
  author={Pelecanos, Jason and Wang, Quan and Huang, Yiling and Moreno, Ignacio Lopez},
  booktitle={The Speaker and Language Recognition Workshop (Odyssey)},
  year={2022}
}