Layer-aware TDNN: Speaker Recognition Using Multi-Layer Features from Pre-Trained Models (2026 ICAIIC)

We share a PyTorch implementation of our experiments here.

The figure above overviews the proposed framework (a & c) to pool speaker embedding from the multi-layered nature of pre-trained models.

Environment Supports & Python Requirements

• We recommend you to visit Previous Versions (v1.12.0) for PyTorch installation including torchaudio==0.12.0.
  [2025.12.17 update] A new code implementation is powered by PyTorch v2.6.0.

Use the requirements.txt to install the rest of the Python dependencies.
Ubuntu-Soundfile and conda-ffmpeg packages would be required for downloading and preprocessing data, and you can install them as below.

$ pip install -r requirements.txt
$ apt-get install python3-soundfile
$ conda install -c conda-forge ffmpeg

1. Dataset Preparation

The datasets can be downloaded from here:

• VCTK CSTR Corpus

• LibriSpeech

• VoxCeleb 1 & 2
  We use clovaai/voxceleb_trainer to download VoxCeleb datasets, released under the MIT licence.
  Follow the data preparation script, till you convert the VoxCeleb 2 audio format aac(.m4a) into .wav file.

2. Data Preprocessing & Evaluation Split

The following scripts are to preprocess audio data and build evaluation trials from each dataset.

• However, you can skip the "# set split" part, since we have uploaded the ready-made splits and trials used for our experiments in the file tree.
  Please check the contents in data/VCTK-Corpus/, data/LibriSpeech/, data/VoxCeleb/ [;speakers/, ;trials/] first.

VCTK CSTR Corpus

# preprocessing
$ python ./src/preprocess/process-VCTK.py --read_path SRC_PATH

Remove speaker [p280, p315] of risen technical issues.
Drop samples (no.000~no.024), where the same transcript is used for each number.
Resample audio sources to meet the sample rate in common (48K → 16K).

# set split
$ python ./src/preprocess/split-VCTK-0-speakers.py
$ python ./src/preprocess/split-VCTK-1-rawtrials.py
$ python ./src/preprocess/split-VCTK-2-balancedtrials.py

Subset the total speaker pool into train, validation, and test speaker subsets.
Check the match of speaker meta-info (Gender | Age | Accents | Region | Label) given the total combination.
Sample the trials with a balance to the label distribution and meta-info matches.

LibriSpeech

# preprocessing
$ python ./src/preprocess/process-LibriSpeech.py --read_path SRC_PATH

Convert audio format .flac to .wav file.

# set split
$ python ./src/preprocess/split-LibriSpeech-1-rawtrials.py
$ python ./src/preprocess/split-LibriSpeech-2-balancedtrials.py

Check the match of speaker meta-info (Gender(SEX) | Label) given the total combination of samples.
Sample the trials with a balance to the label distribution and meta-info matches.

VoxCeleb 1 & 2

$ mv ./data/VoxCeleb/*_wav/ ./data/VoxCeleb/preprocess/

No special data preprocessing required.

# set split
$ python ./src/preprocess/split-VoxCeleb-0-speakers.py
$ python ./src/preprocess/split-VoxCeleb-2-balancedtrials.py

List up the speakers in each subsets, and convert 'Vox1-O' evaluation path file format .txt to .csv.
Sample the trials with a balance to the label distribution.

3. Run Experiments

Loggings, weights, and training configurations will be saved under res/ directory.
The result folder will be created in local-YYYYMMDD-HHmmss format by default.

To use neptune.ai logging, set your configuration in src/configs/neptune/neptune-logger-config.yaml and add --neptune in command line.
The experiment ID created at your neptune.ai [project] will be the name of the output directory.

• General usage examples,

# Running directly through the command line
$ CUDA_VISIBLE_DEVICES=0 python ./src/main.py train VCTK UniPool --use_pretrain --frz_pretrain --batch_size 128 --seed 9973 --backbone_cfg facebook/wav2vec2-base --nb_total_step 10000 --nb_steps_eval 1000;

# Or you can use a shell file for the multiple commands.
$ ./src/run.sh

• Adjusting hyperparameters directly by command.

$ python ./src/main.py -h
usage: main.py [1-action] [2-data] [3-model folder] [-h]

positional arguments (required):
  [1] action:  {train,eval}
  [2] data  :  {VCTK,LibriSpeech,Vox1,Vox2}
  [3] model :  {X-vector,ECAPA-TDNN,SincNet,ExploreWV2,FinetuneWV2,L-TDNN_ECAPA}

optional arguments in general:
  -h, --help                 show this help message and exit
  --quickrun                 quick check for the running experiment on the modification, set as True if given
  --skiptest                 skip evaluation for the testset during training, set as True if given
  --neptune                  log experiment with neptune logger, set as True if given
  --workers     WORKERS      the number of cpu-cores to use from dataloader (per each device), defaults to: 4
  --device     [DEVICE0,]    specify the list of index numbers to use the cuda devices, defaults to: [0]
  --seed        SEED         integer value of random number initiation, defaults to: 42
  --eval_path   EVAL_PATH    result path to load model on the "action" given as {eval}, defaults to: None
  --description DESCRIPTION  user parameter for specifying certain version, Defaults to "Untitled".

keyword arguments:
  --kwarg KWARG              dynamically modifies any of the hyperparameters declared in ../configs/.../...*.yaml or ./benchmarks/...*.yaml
  (e.g.) --lr 0.001 --batch_size 64 --nb_total_step 25000 ...

4. Evaluate

• The following command line will conduct the test evaluation with the best-validated model parameter from the configuration of DIR_NAME

CUDA_VISIBLE_DEVICES=0 python ./src/main.py eval _ _ --eval_path DIR_NAME;

• You can also conduct the cross-dataset evaluation by modifying the command like this.

CUDA_VISIBLE_DEVICES=0 python ./src/main.py eval Vox1 _ --eval_path DIR_NAME;

Citation

@article{kim2024layer,
  title={Layer-aware TDNN: Speaker Recognition Using Multi-Layer Features from Pre-Trained Models},
  author={Kim, Jin Sob and Park, Hyun Joon and Shin, Wooseok and Yun, Juan and Han, Sung Won},
  journal={arXiv preprint arXiv.2409.07770v2},
  year={2024}
}

@article{kim2024universal,
  title={Universal Pooling Method of Multi-layer Features from Pretrained Models for Speaker Verification},
  author={Kim, Jin Sob and Park, Hyun Joon and Shin, Wooseok and Han, Sung Won},
  journal={arXiv preprint arXiv.2409.07770},
  year={2024}
}

License

This repository is released under the MIT license.

As we commented from each comparison src/benchmarks/ models, the projects below are referred to reproduce the implementations.
"Official" means that the author of the paper released the project.

Open to the public

• X-vector/: cvqluu/TDNN

Released under MIT license

• SincNet/ : mravanelli/SincNet (Official)
• ECAPA-TDNN/ : TaoRuijie/ECAPA-TDNN
• FinetuneWV2/: nikvaessen/w2v2-speaker (Official)

Some of the src/utils/ are quoted from below.

• metrics.py: speechbrain/utils/metric_stats.py
• sampler.py: SeungjunNah/DeepDeblur-PyTorch/src/data/sampler.py