FlaxSpeaker: cn_ecapa_tdnn_small_arcface (ecapa_tdnn - 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 (1,800 steps / ~19.80 epochs on 2,909 utterances from 119 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 1,800 gradient steps (S_safe = X + 2 epochs safety margin, warmup 180 steps) with mini-batches of 32 utterances (8 speakers x 4 utterances), i.e. ~19.80 epochs over the 2,909-utterance CN-Celeb2 training pool (119 speakers), on a CPU workstation with seed 42. Angular margin warmup is linearly ramped over the first 250 steps (margin_warmup_steps) before holding the target margin. Across the training trajectory, in-domain CN-Celeb2 EER progressed from step 400: 24.73% (1-utt) / 23.60% (3-utt) -> step 1600 (knee): 19.03% (1-utt) / 14.30% (3-utt) -> step 1800 (final): 18.67% (1-utt) / 13.80% (3-utt) (Universal Pooled final: 18.98% 1-utt / 14.80% 3-utt). While training loss and validation EER have reached their small-corpus convergence plateau on this 119-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 6.7785; current code with fixed 180-frame crops 7.5219, with U(90, 180) crops 6.7785 (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	cn_ecapa_tdnn_small_arcface
Hugging Face Repository	wq2012/cn_ecapa_tdnn_small_arcface
Hugging Face Collection	FlaxSpeaker Collection
GitHub Repository	wq2012/FlaxSpeaker
PyPI Package	flaxspeaker
Backbone Architecture	ecapa_tdnn (small)
Temporal Pooling	asp
Training Objective	arcface (arcface)
Verification Scorer	Cosine similarity (cosine)
Output Representation Dim	192
Inference Parameters (encoder + scoring head; excludes training-only classifier weights)	1,684,416
Training Data	CN-Celeb2 only (119 speakers, 2,909 utterances)
Training Steps / Batch	1,800 steps x 32 utterances (~19.80 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 >= 0.4.0 (pip install -U flaxspeaker) for full margin_warmup_steps config compatibility with the v0.4.0 converged checkpoints (>= 0.3.0 introduced load_pretrained; the legacy 0.2.0 wheel cannot load Hub checkpoints).

from flaxspeaker import load_pretrained

model = load_pretrained("wq2012/cn_ecapa_tdnn_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/cn_ecapa_tdnn_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	EER (%)	95% Bootstrap CI (%)	minDCF@0.01	minDCF@0.10	ROC-AUC
LibriSpeech (English Read)	3,000	15.23%	[14.07, 16.58]	0.7760	0.6473	0.9306
CN-Celeb2 (Chinese Multi-Genre)	3,000	18.67%	[17.18, 19.88]	0.8907	0.7113	0.8991
VoxCeleb1 (In-the-Wild Celebrity)	2,000	20.75%	[18.70, 22.40]	0.9800	0.8740	0.8740
Universal Pooled (All 3 Domains)	8,000	18.98%	[18.07, 19.77]	0.9060	0.7712	0.8969
Universal Macro-Average (3 Domains)	—	18.22%	—	0.8822	0.7442	0.9012
VoxCeleb1-O 10-speaker subset (see caveat)	2,500	22.21%	[20.51, 23.99]	0.9866	0.8123	0.8636

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	9.40%	[7.55, 11.20]	0.6980	0.9680
CN-Celeb2 (3-Utt)	1,000	13.80%	[11.60, 15.80]	0.8440	0.9404
VoxCeleb1 (3-Utt)	1,000	17.40%	[14.85, 19.60]	0.9640	0.9061
Universal Pooled (3-Utt)	3,000	14.80%	[13.47, 16.12]	0.9000	0.9351
Universal Macro-Average (3-Utt)	—	13.53%	—	0.8353	0.9382

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: 1,800 gradient steps x 32 utterances (N = 8 speakers x M = 4, ~19.80 epochs) on <= 25 utterances/speaker from 119 speakers (2,909 utterances); the source papers train for ~500k steps with N = 16-64 / M = 8-10 on thousands of speakers.
• Augmentation: synthetic log-Mel-domain noise + one-tap smoothing (specaug.apply_feature_mtr_augmentation), not waveform MTR with recorded noise/RIR corpora.
• ECAPA-TDNN: LayerNorm instead of BatchNorm.
• Frontend: 80-D natural-log Mel + frame-level energy VAD (26 dB dynamic range) + CMVN; evaluation centre-crops to at most 500 frames (5 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}
}