Skip to content

This project partially embodies the state-of-the-art practices in speaker verification technology up until 2020, while attaining the state-of-the-art performance on the VoxCeleb1 test sets.

License

Notifications You must be signed in to change notification settings

xx205/voxsrc2020_speaker_verification

Repository files navigation

Overview

This project partially reflects the best practices in speaker verification technology up till 2020.

For more details, please refer to the link: The VoxSRC Workshop 2020.

Data

  • Training data are prepared following Kaldi voxceleb recipe: VoxCeleb2_dev and its four augmented versions are generated using RIRS_NOISES and MUSAN datasets.

  • 80-dimensional (or 40-dimensional) FBANKs are extracted.

  • Speaker augmentation is not applied.

  • Online data augmentation is not applied.

Models

  • 1D-conv model: 512-d TDNN model following Kaldi voxceleb recipe.

  • 2D-conv models: Res2Net model and DPN (Dual Path Networks) model.

Training

Scoring

  • Cosine similarity scoring.

  • Adaptive symmetric normalization (asnorm) on cosine scores.

  • Quality measure function (QMF) or score calibration is not applied.

Notes

  • Data preparation takes a long time. It is recommended to run data preparation script on a server. With a D48s_v5 Virtual Machine on Azure (48 vCPUs, 192 GB memory), it takes about 10 hours to complete the whole process.

  • For the smallest Res2Net model, res2net50_w8_s6_c16, it takes about 52 + 5 hours to train the model on a single NVIDIA Geforce RTX 4090 GPU.

  • For the largest Res2Net model, res2net200_w8_s6_c16, it takes about 76 + 8 hours to train the model on eight NVIDIA Geforce RTX 3090 GPUs.

Experiments

Environment setup

  • Install TensorFlow 1.x and Kaldi toolkit manually

  • or setup a docker container with Dockerfile (recommended):

    docker build -t xx205/tensorflow:23.03-tf1-py3-kaldi -f Dockerfile .
  • or pull the container directly (recommended):

    docker pull xx205/tensorflow:23.03-tf1-py3-kaldi

Data preparation

  • Run docker container:

    docker run --rm -it --gpus device=all -v `pwd`:`pwd` --ipc=host --ulimit memlock=-1 xx205/tensorflow:23.03-tf1-py3-kaldi
  • Clone this repository:

    git clone https://github.com/xx205/voxsrc2020_speaker_verification
  • Run data preparation script (note, 80-dimensional FBANKs are extracted by default. FBANK dimension can be set in global_config.sh):

    cd voxsrc2020_speaker_verification
    
    bash prepare_data.sh

Model training

  • Go to working directory:

    cd tensorflow
  • Train a Res2Net model with JTBD's VoxSRC2020 setup on VoxCeleb2_dev_aug training set:

    bash run_res2net_local_vox2_dev_aug.sh models.res2net_model res2net50_w24_s4_c32
  • Export trained checkpoint to pb file for inference:

    bash export_inference_model.sh \
        models.res2net_model \
        res2net50_w24_s4_c32 \
        exp/voxceleb2_dev_aug/res2net50_w24_s4_c32_sc_cm_linear_frames200_scale32_margin0.2_8GPUs \
        3 \
        sc_cm_linear/kernel
  • Evaluate trained Res2Net model performance on VoxCeleb1 Test/Extended/Hard trials:

    export saved_model=exp/voxceleb2_dev_aug/res2net50_w24_s4_c32_sc_cm_linear_frames200_scale32_margin0.2_8GPUs_122636.pb
    
    bash eval_inference_model.sh ${saved_model} 3
    
    for partition in T E H; do
        for approach in cosine snorm; do
            python3 eer_minDCF.py \
                --trial ../data/voxceleb1_trials/list_test_${partition}.txt \
                --score ${saved_model%.pb}_embeddings/voxceleb1/${approach}_${partition}.txt
        done
    done
  • Finetune the Res2Net model with LMFT on VoxCeleb2_dev training set:

    bash run_res2net_finetune_local_vox2_dev.sh models.res2net_model res2net50_w24_s4_c32
  • Export finetuned checkpoint to pb file for inference:

    bash export_inference_model.sh \
        models.res2net_model \
        res2net50_w24_s4_c32 \
        exp/voxceleb2_dev/res2net50_w24_s4_c32_sc_cm_linear_frames600_scale32_margin0.4_8GPUs \
        3 \
        sc_cm_linear/kernel
  • Evaluate finetuned Res2Net model performance on VoxCeleb1 Test/Extended/Hard trials:

    export saved_model=exp/voxceleb2_dev/res2net50_w24_s4_c32_sc_cm_linear_frames600_scale32_margin0.4_8GPUs_127968.pb
    
    bash eval_inference_model.sh ${saved_model} 3
    
    for partition in T E H; do
        for approach in cosine snorm; do
            python3 eer_minDCF.py \
                --trial ../data/voxceleb1_trials/list_test_${partition}.txt \
                --score ${saved_model%.pb}_embeddings/voxceleb1/${approach}_${partition}.txt
        done
    done

Results

1. VoxCeleb1_dev_aug as training data, 40-d FBANK features

Model configurations

model loss function subcenter sample length margin # parameters pooling function
tdnn_voxsrc2020 cm_linear_voxsrc2020 320 (0.2, 0.1) 3.5 M Stats Pool
dpn68_voxsrc2020 cm_linear_voxsrc2020 320 (0.2, 0.1) 13.9 M Stats Pool

Results on VoxCeleb1_Test

Vox1_Test
(EER/minDCF0.01)
tdnn_voxsrc2020 (cosine) 3.4398%/0.3339
dpn68_voxsrc2020 (cosine) 2.0894%/0.2544

2. VoxCeleb2_dev_aug as training data, 40-d FBANK features

Model configurations

model loss function subcenter sample length margin # parameters pooling function
dpn68_voxsrc2020 cm_linear_voxsrc2020 320 (0.2, 0.1) 13.9 M Stats Pool
res2net50_w24_s4_c64_voxsrc2020 cm_linear_voxsrc2020 320 (0.2, 0.1) 26.9 M Stats Pool

Results on VoxCeleb1_Test/Extended/Hard

Vox1_Test
(EER/minDCF0.01)
Extended
(EER/minDCF0.01)
Hard
(EER/minDCF0.01)
dpn68_voxsrc2020 (cosine) 0.9517%/0.0884 1.0217%/0.1125 1.8649%/0.1833
dpn68_voxsrc2020 (asnorm) 1 0.8347%/0.0879 0.9452%/0.0996 1.6401%/0.1527
res2net50_w24_s4_c64_voxsrc2020 (cosine) 0.9942%/0.1163 0.9924%/0.1152 1.8387%/0.1857
res2net50_w24_s4_c64_voxsrc2020 (asnorm) 2 0.8400%/0.0931 0.9051%/0.0970 1.5976%/0.1471

3. VoxCeleb2_dev_aug as training data, 40-d FBANK features, LMFT on VoxCeleb2_dev

Model configurations

model loss function subcenter sample length margin # parameters pooling function
dpn68 sc_cm_linear 200 0.2 13.9 M Stats Pool
dpn68+LMFT sc_cm_linear 600 0.4 13.9 M Stats Pool
res2net50_w24_s4_c64 sc_cm_linear 200 0.2 26.9 M Stats Pool
res2net50_w24_s4_c64+LMFT sc_cm_linear 600 0.4 26.9 M Stats Pool

Results on VoxCeleb1_Test/Extended/Hard

Vox1_Test
(EER/minDCF0.01)
Extended
(EER/minDCF0.01)
Hard
(EER/minDCF0.01)
dpn68 (cosine) 1.0952%/0.1052 1.1725%/0.1322 2.0998%/0.2053
dpn68 (asnorm) 0.9783%/0.1081 1.0707%/0.1184 1.8656%/0.1742
dpn68+LMFT (cosine) 0.8666%/0.0824 0.9634%/0.1058 1.6910%/0.1641
dpn68+LMFT (asnorm) 0.8081%/0.0735 0.8934%/0.0925 1.5076%/0.1414
res2net50_w24_s4_c64 (cosine) 1.0580%/0.1089 1.0728%/0.1226 1.9778%/0.1911
res2net50_w24_s4_c64 (asnorm) 0.9038%/0.0989 0.9824%/0.1086 1.7269%/0.1621
res2net50_w24_s4_c64+LMFT (cosine) 0.9198%/0.1005 0.9258%/0.1006 1.6957%/0.1650
res2net50_w24_s4_c64+LMFT (asnorm) 0.7922%/0.0804 0.8482%/0.0903 1.4854%/0.1353

4. VoxCeleb2_dev_aug as training data, 80-d FBANK features, LMFT on VoxCeleb2_dev

Model configurations

model loss function subcenter sample length margin # parameters pooling function
res2net50_w24_s4_c64 sc_cm_linear 200 0.2 32.2 M Stats Pool
res2net50_w24_s4_c64+LMFT sc_cm_linear 600 0.4 32.2 M Stats Pool
res2net50_w24_s4_c32 sc_cm_linear 200 0.2 17.7 M Stats Pool
res2net50_w24_s4_c32+LMFT sc_cm_linear 600 0.4 17.7 M Stats Pool
res2net50_w8_s6_c16 sc_cm_linear 200 0.2 4.8 M Stats Pool
res2net50_w8_s6_c16+LMFT sc_cm_linear 600 0.4 4.8 M Stats Pool
res2net101_w24_s4_c32_att sc_cm_linear 200 0.2 29.3 M Att Stats Pool
res2net101_w24_s4_c32_att+LMFT sc_cm_linear 600 0.4 29.3 M Att Stats Pool
res2net152_w24_s4_c32_att sc_cm_linear 200 0.2 32.9 M Att Stats Pool
res2net152_w24_s4_c32_att+LMFT sc_cm_linear 600 0.4 32.9 M Att Stats Pool
res2net200_w24_s4_c32_att sc_cm_linear 200 0.2 35.5 M Att Stats Pool
res2net200_w24_s4_c32_att+LMFT sc_cm_linear 600 0.4 35.5 M Att Stats Pool

Results on VoxCeleb1_Test/Extended/Hard, VoxSRC2022_dev

Vox1_Test
(EER/minDCF0.01)
Extended
(EER/minDCF0.01)
Hard
(EER/minDCF0.01)
VoxSRC2022_dev
(EER/minDCF0.05)
res2net50_w24_s4_c64 (cosine) 0.9304%/0.0855 1.0345%/0.1109 1.8228%/0.1777
res2net50_w24_s4_c64 (asnorm) 0.7656%/0.0662 0.8989%/0.0964 1.5381%/0.1423
res2net50_w24_s4_c64+LMFT (cosine) 0.7762%/0.0702 0.8589%/0.0933 1.5236%/0.1506
res2net50_w24_s4_c64+LMFT (asnorm) 0.6805%/0.0579 0.8024%/0.0827 1.3703%/0.1210
res2net50_w24_s4_c32 (cosine) 0.8347%/0.0771 0.9893%/0.1104 1.8242%/0.1794
res2net50_w24_s4_c32 (asnorm) 0.7018%/0.0673 0.9062%/0.0986 1.5860%/0.1466
res2net50_w24_s4_c32+LMFT (cosine) 0.6327%/0.0622 0.8500%/0.0937 1.5352%/0.1549
res2net50_w24_s4_c32+LMFT (asnorm) 0.5529%/0.0625 0.8003%/0.0820 1.3751%/0.1271
res2net50_w8_s6_c16 (cosine) 1.0474%/0.1133 1.1721%/0.1288 2.0929%/0.2074
res2net50_w8_s6_c16 (asnorm) 0.9889%/0.1024 1.0659%/0.1164 1.8627%/0.1777
res2net50_w8_s6_c16+LMFT (cosine) 0.8666%/0.0839 0.9772%/0.1019 1.7360%/0.1678
res2net50_w8_s6_c16+LMFT (asnorm) 0.7869%/0.0821 0.8903%/0.0917 1.5370%/0.1414
res2net101_w24_s4_c32_att (cosine) 0.6539%/0.0655 0.7837%/0.0823 1.4125%/0.1395
res2net101_w24_s4_c32_att (asnorm) 0.5742%/0.0664 0.7230%/0.0749 1.2679%/0.1187
res2net101_w24_s4_c32_att+LMFT (cosine) 0.5795%/0.0512 0.6526%/0.0677 1.2099%/0.1174
res2net101_w24_s4_c32_att+LMFT (asnorm) 0.5210%/0.0534 0.6081%/0.0618 1.0940%/0.0999
res2net152_w24_s4_c32_att (cosine) 0.5476%/0.0507 0.7313%/0.0811 1.3453%/0.1331
res2net152_w24_s4_c32_att (asnorm) 0.4891%/0.0530 0.6754%/0.0729 1.1957%/0.1101
res2net152_w24_s4_c32_att+LMFT (cosine) 0.4732%/0.0425 0.6516%/0.0676 1.1957%/0.1119
res2net152_w24_s4_c32_att+LMFT (asnorm) 0.4572%/0.0463 0.6099%/0.0587 1.0737%/0.0965
res2net200_w24_s4_c32_att (cosine) 0.4944%/0.0418 0.7137%/0.0780 1.2897%/0.1267
res2net200_w24_s4_c32_att (asnorm) 0.4200%/0.0501 0.6658%/0.0690 1.1496%/0.1019
res2net200_w24_s4_c32_att+LMFT (cosine) 0.4041%/0.0390 0.6330%/0.0672 1.1641%/0.1113
res2net200_w24_s4_c32_att+LMFT (asnorm) 0.3668%/0.0388 0.5930%/0.0581 1.0330%/0.0912 1.5017/0.0974

Footnotes

  1. DPN68 model with B+α+1 configuration (https://arxiv.org/pdf/2011.00200.pdf)

  2. Res2Net50 model with B+α+5 configuration (https://arxiv.org/pdf/2011.00200.pdf)

About

This project partially embodies the state-of-the-art practices in speaker verification technology up until 2020, while attaining the state-of-the-art performance on the VoxCeleb1 test sets.

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published