MR-RawNet

This repository contains official pytorch implementation and pre-trained models for following paper:

• Title : MR-RawNet: Speaker verification system with multiple temporal resolutions for variable duration utterances using raw waveforms
• Autor : Seung-bin Kim, Chan-yeong Lim, Jungwoo Heo, Ju-ho Kim, Hyun-seo Shin, Kyo-Won Koo and Ha-Jin Yu
• Comments : this paper has been accepted at Interspeech 2024 and is available here.

Abstract

In speaker verification systems, the utilization of short utterances presents a persistent challenge, leading to performance degradation primarily due to insufficient phonetic information to characterize the speakers. To overcome this obstacle, we propose a novel structure, MR-RawNet, designed to enhance the robustness of speaker verification systems against variable duration utterances using raw waveforms. The MR-RawNet extracts time-frequency representations from raw waveforms via a multi-resolution feature extractor that optimally adjusts both temporal and spectral resolutions simultaneously. Furthermore, we apply a multi-resolution attention block that focuses on diverse and extensive temporal contexts, ensuring robustness against changes in utterance length. The experimental results, conducted on VoxCeleb1 dataset, demonstrate that the MR-RawNet exhibits superior performance in handling utterances of variable duration compared to other raw waveform-based systems.

Our experimental code was modified based on voxceleb_trainer, and we referenced the baseline code at here.

Data

The VoxCeleb datasets were used for training and test.

The train list should contain the identity and the file path, one line per utterance, as follows:

id00000 id00000/youtube_key/12345.wav
id00012 id00012/21Uxsk56VDQ/00001.wav

The train list for VoxCeleb2 can be download from here, and the test lists for VoxCeleb1 can be downloaded from here.

For data augmentation, the following script can be used to download and prepare.

python3 ./dataprep.py --save_path data --augment

We also performed an out-of-domain evaluation using the VOiCES development set.

Each dataset must be downloaded in advance for training and testing, and its path must be mapped to the docker environment.

Environment

Docker image (nvcr.io/nvidia/pytorch:23.07-py3) of Nvidia GPU Cloud was used for conducting our experiments.

Make docker image and activate docker container.

./docker/build.sh
./docker/run.sh

Note that you need to modify the mapping path before running the 'run.sh' file.

Training

• MR-RawNet on a single GPU

python3 ./trainSpeakerNet.py --config ./configs/MR_RawNet.yaml

• MR-RawNet on multiple GPUs

CUDA_VISIBLE_DEVICES=0,1 python3 ./trainSpeakerNet.py --config ./configs/MR_RawNet.yaml --distributed

Use --distributed flag to enable distributed training. If you are running more than one distributed training session, you need to change the --port argument.

Note that the configuration file overrides the arguments passed via command line.

Test

The following script should return: EER 0.8294.

python3 ./trainSpeakerNet.py --eval --config ./configs/MR_RawNet.yaml --initial_model MR_RawNet.pt

Citation

Please cite if you make use of the code.

@article{kim2024mrrawnet,
  title={MR-RawNet: Speaker verification system with multiple temporal resolutions for variable duration utterances using raw waveforms},
  author={Kim, Seung-bin and Lim, Chan-yeong and Heo, Jungwoo and Kim, Ju-ho and Shin, Hyun-seo and Koo, Kyo-Won and Yu, Ha-Jin},
  journal={arXiv preprint arXiv:2406.07103},
  year={2024}
}

@inproceedings{kim2024mrrawnet,
  title={MR-RawNet: Speaker verification system with multiple temporal resolutions for variable duration utterances using raw waveforms},
  author={Kim, Seung-bin and Lim, Chan-yeong and Heo, Jungwoo and Kim, Ju-ho and Shin, Hyun-seo and Koo, Kyo-Won and Yu, Ha-Jin},
  booktitle={Proc. Interspeech},
  year={2024}
}