HAM-Softmax: Hyperbolic Additive Margin Softmax with Hierarchical Information for Speaker Verification

Official implementation of Hyperbolic Additive Margin Softmax with Hierarchical Information for Speaker Verification, accepted by ICASSP 2026.

Data Preparation

Prepare VoxCeleb or CN-Celeb style audio directories, then replace the dataset roots in utils.py with your local paths:

vox1_dev_path = '/path/to/voxceleb1/dev/wav/'
vox2_dev_path = '/path/to/voxceleb2/dev/aac/'
cn_dev_path = '/path/to/cnceleb'
eval_path = '/path/to/voxceleb1'

The repository uses list files under data:

data/
  v1_clean.txt
  v2_clean.txt
  cn_clean.txt
  vox_O.txt
  vox_E.txt
  vox_H.txt
  vox_EH_list.txt
  CN.Eval_list.txt

Training list format:

speaker_id relative/path/to/audio.wav

VoxCeleb-O/E/H evaluation list format:

label enroll_utterance test_utterance

For CN-Celeb evaluation, replace the corresponding paths in utils.py:

eval_list1 = '/path/to/cn_veri_test.txt'
eval_list2 = '/path/to/cn_veri_test.txt'
eval_path = '/path/to/cnceleb/cn_1/eval'

For augmentation, replace the MUSAN and RIR paths in datasets.py:

self.musan_path = "/path/to/musan"
self.rir_path = "/path/to/RIRS_NOISES/simulated_rirs"

Training

General training command:

CUDA_VISIBLE_DEVICES=0 python main.py \
  --dataset v2 \
  --lambda_1 0 \
  --lambda_2 1 \
  --h_C 3 \
  --h_s 30 \
  --h_m 0.2 \
  --batch_size 256 \
  --augment

Important arguments:

Argument	Description
--dataset	Dataset key: v1, v2, or cn.
--loss_type	Euclidean auxiliary loss: ce, ces, am, aam, or ram.
--lambda_1	Weight of the Euclidean loss branch.
--lambda_2	Weight of the hyperbolic loss branch.
--h_C	Hyperbolic curvature parameter.
--h_m	Hyperbolic additive margin. Use 0 for H-Softmax.
--h_s	Hyperbolic scale factor.
--augment	Enable waveform augmentation.

Example Experiments

The provided run.sh contains three example VoxCeleb1 runs:

./run.sh

They correspond to:

Experiment	Key setting
H-Softmax	h_C=5, h_m=0, h_s=30, lambda_1=0, lambda_2=1
HAM-Softmax	h_C=3, h_m=0.2, h_s=30, lambda_1=0, lambda_2=1
E.H. AM-Softmax	loss_type=ram, h_C=3, h_m=0.2, h_s=30, lambda_1=0.3, lambda_2=0.7

Evaluation and Outputs

During training, the script evaluates periodically according to the dataset setting in utils.py. Results and checkpoints are saved under:

exps/<experiment_name>/
  train.log
  result_<experiment_name>.csv
  Vox-EH.csv
  model/

When training on VoxCeleb2 (--dataset v2), the final model is additionally evaluated on VoxCeleb-E and VoxCeleb-H with the fast evaluation dataset. The results are saved to Vox-EH.csv.

Citation

If this repository is helpful for your research, please cite:

@INPROCEEDINGS{11463316,
  author={Fang, Zhihua and He, Liang},
  booktitle={ICASSP 2026 - 2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  title={Hyperbolic Additive Margin Softmax with Hierarchical Information for Speaker Verification},
  year={2026},
  pages={19017-19021},
  doi={10.1109/ICASSP55912.2026.11463316}
}