Unofficial implementation of "Speaker recognition from raw waveform with sincnet" paper.

Usage

Google Colab

To run code in google colab, open Experiments report.ipynb, run the first cell to clone the repository, then add TRAIN and TEST folders from timit dataset to the datasets/timit/ folder and run ./bins/timit_lowercase.sh script to convert all folders and files names to lower case (you will see warning messages like this mv: cannot move 'datasets/timit/train' to a subdirectory of itself, 'datasets/timit/train/train' but it's fine).

If you want to evaluate pretrained model on speaker verification task, you will need to add them in src/model folder with names cnn.pt and sinc.pt for baseline and SincNet respectively. Then, from the src folder run this command to compute d-vecots
!python speaker_verification/compute_d_vectors.py model_type pretreained_path compute_split --save_to d_vectors_save_path.npy
you need to compute d-vectors for both sv.scp and train.scp splits and then you can compute err via
!python speaker_verification/speaker_verification.py model_type pretrained path d_vectors_speakers d_vector_imposters.
For more details you can look at the soucre code of compute_d_vectors.py and speaker_verification.py or how they are used in notebook.

If you want to train from scratch, run main.py. If you want to train baseline model change model.type in configs/cfg.yaml' to cnn`.

Docker

To run code in docker container:

• run ./bins/build_image.sh and ./bins/run_container.sh scripts
• copy TRAIN and TEST folders from timit dataset to the datasets/timit folder inside the container via runnig
  docker cp path/to/TRAIN/ cuda0:/home/sincnet/src/datasets/timit/
• run docker exec -i -it cuda0 bash -c "./bins/timit_lowercase.sh" to cast all folders and files in dataset to lower case.

To train model from scratch run docker exec -i -it cuda0 bash -c "cd src && python main.py".

To evaluate pretrained model on speaker verification task:

• copy pretrained weights in src/model/
• compute d-vecots, do it via
  docker exec -i -it cuda0 bash -c "cd src && python speaker_verification/compute_d_vectors.py sinc model/sinc.pt train.scp --save_to d_vectors_sinc.npy"
docker exec -i -it cuda0 bash -c "cd src && python speaker_verification/compute_d_vectors.py sinc model/sinc.pt sv.scp --save_to d_vectors_sinc_unseen.npy"
• compute err via
  docker exec -i -it cuda0 bash -c "cd src && python speaker_verification/speaker_verification.py sinc model/sinc.pt d_vectors_sinc.npy d_vectors_sinc_unseen.npy type" (but first change type argument to cos or softmax)

Weights

Baseline: link.
SincNet: link.