FlaxSpeaker: univ_full_eres2conformer_large (eres2conformer - large)

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 original papers were run on Google-internal infrastructure and data; this checkpoint is a full-data open-corpus universal flagship trained on 12,193 speakers (972,254 clean utterances + MUSAN/RIR/speed augmentation) across VoxCeleb2-dev, VoxCeleb1-dev, CN-Celeb 1+2, and LibriSpeech 960h with zero speaker overlap against official evaluation sets.

Training Budget (read this first)

This flagship model was trained for 93,000 gradient steps (warmup 2000 steps, batch size 512) over the full multi-corpus universal training pool (VoxCeleb1/2 + CN-Celeb1/2 + LibriSpeech 960h) (12,193 base speakers, 972,254 clean utterances, expanded via 3-way speed perturbation + MUSAN/RIR).

Training-crop provenance. This checkpoint was trained with U(140, 200)-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 11.7771; current code with fixed 200-frame crops 11.7526, with U(140, 200) crops 11.781 (match: variable_approx). The variable-crop CPU float32 probe is within 0.0039 of the archived TPU bfloat16 step-0 loss, whereas the fixed-crop arm is more than 5x farther away. 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 %.

Model Summary

Property Value
Model ID univ_full_eres2conformer_large
Hugging Face Repository wq2012/univ_full_eres2conformer_large
Hugging Face Collection FlaxSpeaker Collection
GitHub Repository wq2012/FlaxSpeaker
PyPI Package flaxspeaker
Backbone Architecture eres2conformer (large)
Temporal Pooling catp
Training Objective Sub-Center ArcFace (Deng et al. 2020; K=3, scale=32, margin=0.3) (subcenter_arcface)
Verification Scorer Cosine similarity (cosine)
Output Representation Dim 256
Inference Parameters (encoder + scoring head; excludes training-only classifier weights) 31,834,352
Training Data VoxCeleb2-dev + VoxCeleb1-dev + CN-Celeb1/2 + LibriSpeech 960h (full multi-domain SOTA flagship, 12,193 speakers, 972,254 utterances)
Training Steps / Batch 93,000 steps x 512 utterances (~46.87 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)
  • model_fp32.tflite / model_int8.tflite: jax2tf exports of the encoder only with a fixed, unmasked 180-frame (1.8 s) input window [1, 180, 80] -> [1, 256]. All EERs in the main tables below come from full-sequence masked JAX inference (up to 3000 frames / 30 s); the TFLite models see a centre crop and the reported cosine fidelity (>= 0.9999 vs. JAX) was measured on 180-frame inputs.
    • Measured 180-frame TFLite 1-utt EER (VoxCeleb1-O official (37,611 trials, 180-frame center crop)): FP32 2.38%, INT8 2.41%, mean cos(FP32, INT8) = 0.999707 (vs. full-sequence masked JAX: 0.84%)

Quickstart Usage

Python API (from Hugging Face Hub or Local Directory)

Requires flaxspeaker >= 0.7.0 (pip install -U flaxspeaker).

from flaxspeaker import load_pretrained

model = load_pretrained("wq2012/univ_full_eres2conformer_large")
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_full_eres2conformer_large \
    --enroll_audio path/to/enroll.flac --test_audio path/to/test.flac

Official Zero-Shot Speaker Verification Benchmarks

Benchmark Trials Raw EER (%) AS-Norm EER (%) AS-Norm + QMF EER (%) minDCF@0.01 ROC-AUC
VoxCeleb1-O (veri_test2.txt, 40 spks) 37,611 0.84% 0.85% 0.71% 0.0755 0.9997
CN-Celeb-E (eval/lists/trials.lst, 200 spks) 3,484,292 7.83% 7.65% 7.45% 0.3884 0.9752
CN-Celeb-E Multi-Enroll (enroll.map) 3,484,292 7.08% 6.72% β€” 0.3541 0.9794
LibriSpeech test-clean (40 spks) 10,000 0.72% 0.66% 0.65% 0.0222 0.9996
LibriSpeech test-clean (3-Utt) 3,000 0.27% 0.40% β€” 0.0060 1.0000

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 0.90% 0.87% [0.60, 1.33] 0.0267 0.0267 0.9992
CN-Celeb2 (Chinese Multi-Genre) 3,000 8.17% 8.93% [7.22, 9.23] 0.2853 0.2533 0.9738
VoxCeleb1 (In-the-Wild Celebrity) 2,000 0.55% 0.40% [0.25, 0.90] 0.0400 0.0240 0.9999
Universal Pooled (All 3 Domains) 8,000 4.30% 4.58% [3.80, 4.79] 0.1398 0.1248 0.9904
Universal Macro-Average (3 Domains) β€” 3.21% 3.40% β€” 0.1173 0.1013 0.9910

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 0.40% [0.00, 0.90] 0.0040 0.9995
CN-Celeb2 (3-Utt) 1,000 5.10% [3.60, 6.20] 0.1540 0.9886
VoxCeleb1 (3-Utt) 1,000 0.00% [0.00, 0.00] 0.0000 1.0000
Universal Pooled (3-Utt) 3,000 2.30% [1.83, 2.95] 0.0567 0.9966
Universal Macro-Average (3-Utt) β€” 1.83% β€” 0.0527 0.9960

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: 93,000 gradient steps x 512 utterances (~46.87 epochs) across 12,193 base speakers (972,254 clean utterances + 3x speed perturbation + MUSAN/RIR).
  • 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.
  • SimAM-ERes2Conformer: unified 2D multi-scale SimAM-SERes2Net (base_channels=64, num_blocks=(3, 4, 6, 3), scale=4, width_mult=2, SimAM+SE+AFF, temporal_strides=(1, 1, 2, 2)) + Stage-3/4 MFA + 2x 512-D RoPE+QK-Norm+BatchNorm Conformer blocks + catp + BN-Neck.
  • 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}
}
Downloads last month
32
Safetensors
Model size
31.9M params
Tensor type
F32
Β·
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Dataset used to train wq2012/univ_full_eres2conformer_large

Collection including wq2012/univ_full_eres2conformer_large

Papers for wq2012/univ_full_eres2conformer_large