AM-SincNet is a new approach for speaker recognition problems which is based in the neural network architecture SincNet and the additive margin softmax (AM-Softmax) loss function. It uses the architecture of the SincNet, but with an improved AM-Softmax layer.
This repository releases an example of code to perform a speaker recognition experiment on the TIMIT dataset. To run it with other datasets you can have a look at the instructions on the original SincNet repository (https://github.com/mravanelli/SincNet).
We should thank @mravanelli for the SincNet implementation.
For running this experiment we used a Linux environment with Python 3.6.
You can see a list of python dependencies at requirements.txt.
To install it on conda virtual environment (conda install --file requirements.txt).
To install it on pip virtual environment (pip install -r requirements.txt).
To run it on TIMIT dataset we have first to pre-process the data, removing the start and ending silences moments and also normalizing the audio sentences.
python TIMIT_preparation.py $TIMIT_FOLDER $OUTPUT_FOLDER data_lists/TIMIT_all.scp
where:
then, we can run the experiment itself by typing.
python speaker_id.py --cfg=cfg/$CFG_FILE
where:
We have made avaliable several cfg configuration files for the experiments, if you want to run the experiment with the traditional SincNet (with no use of the improved AM-Softmax layer) you must use the SincNet_TIMIT.cfg file, otherwise you can use the SincNet_TIMIT_m0XX.cfg file where the XX denotes the size of the margin parameter that will be used for the AM-Softmax layer.
When training have a look at the cfg configuration file, the output paths for the model and the result (res.res) files are placed there.
We have also made available some results from our experiments, you can check them at exp folder. The resume of the results are saved in the res.res files.
In this repository, we used the TIMIT dataset as a tutorial to show how SincNet works. With the current version of the code, you can easily use a different corpus. To do it you should provide in input the corpora-specific input files (in wav format) and your own labels. You should thus modify the paths into the *.scp files you find in the data_lists folder.
To assign to each sentence the right label, you also have to modify the dictionary "TIMIT_labels.npy". The labels are specified within a python dictionary that contains sentence ids as keys (e.g., "si1027") and speaker_ids as values. Each speaker_id is an integer, ranging from 0 to N_spks-1. In the TIMIT dataset, you can easily retrieve the speaker id from the path (e.g., train/dr1/fcjf0/si1027.wav is the sentence_id "si1027" uttered by the speaker "fcjf0"). For other datasets, you should be able to retrieve in such a way this dictionary containing pairs of speakers and sentence ids.
You should then modify the config file (cfg/SincNet_TIMIT.cfg) according to your new paths. Remember also to change the field "class_lay=462" according to the number of speakers N_spks you have in your dataset.
If you use this code or part of it, please cite us!
@INPROCEEDINGS{8852112,
author={J. A. {Chagas Nunes} and D. {Macêdo} and C. {Zanchettin}},
booktitle={2019 International Joint Conference on Neural Networks (IJCNN)},
title={Additive Margin SincNet for Speaker Recognition},
year={2019},
volume={},
number={},
pages={1-5},
keywords={},
doi={10.1109/IJCNN.2019.8852112},
ISSN={},
month={July},}You can also find the paper at IEEE or the preprint at arXiv.