/DNS-Challenge

This repo contains the scripts, models, and required files for the Deep Noise Suppression (DNS) Challenge.

Primary LanguagePythonCreative Commons Attribution 4.0 InternationalCC-BY-4.0

Deep Noise Suppression (DNS) Challenge 4 - ICASSP 2022

In this repository

This repository contains the datasets and scripts required for ICASSP 2022 DNS Challenge, AKA DNS Challenge 4, or simply DNS4. For more details about the challenge, please see our website and paper. For more details on the testing framework, please visit P.835.

Details

  • The datasets_fullband folder is a placeholder for the datasets. That is, our data downloader script by default will place the downloaded audio data there. After the download, it will contain clean speech, noise, and room impulse responses required for creating the training data for wideband scenario. The script will also download here the test set that participants can use during the development stages.
  • The NSNet2-baseline directory contains the inference scripts and the ONNX model for the baseline Speech Enhancement method for wideband.
  • download-dns-challenge-4.sh - this is the script to download the data. By default, the data will be placed into the ./datasets_fullband/ folder. Please take a look at the script and uncomment the perferred download method. Unmodified, the script performs a dry run and retrieves only the HTTP headers for each archive.
  • download-dns-challenge-4-pdns.sh - Same as above, but for the Personalized DNS Challenge track.
  • noisyspeech_synthesizer_singleprocess.py - is used to synthesize noisy-clean speech pairs for training purposes.
  • noisyspeech_synthesizer.cfg - is the configuration file used to synthesize the data. Users are required to accurately specify different parameters and provide the right paths to the datasets required to synthesize noisy speech.
  • audiolib.py - contains modules required to synthesize datasets.
  • utils.py - contains some utility functions required to synthesize the data.
  • unit_tests_synthesizer.py - contains the unit tests to ensure sanity of the data.
  • requirements.txt - contains all the libraries required for synthesizing the data.

Datasets

DEV Testset: BLIND testset: released at https://dns4public.blob.core.windows.net/dns4archive/blind_testset_bothtracks.zip It has two folders, enrollment_speech (178 clips) and testclips (859 clips). It consists of 638 mobile testclips and rest are desktop/laptop testclips. Each testclips is 10s duration. Both tracks have same testclips. For Track-1 non-personalized DNS, you do not need the enrollment_speech.

Baseline Enhanced clips

Baseline Personalized DEV enhanced clips: https://dns4public.blob.core.windows.net/dns4archive/Baseline_personalized_dev_testset.zip Non-personalized Baseline model: https://github.com/microsoft/DNS-Challenge/tree/master/NSNet2-baseline

Baseline Personalized BLIND enhanced clips: TBD

WAcc script

https://github.com/microsoft/DNS-Challenge/tree/master/WAcc

Wacc transcript

Dev testset: https://github.com/microsoft/DNS-Challenge/blob/master/WAcc/DNSChallenge4_devtest.tsv Blind testset: TBD

Real-time DNS track

The default directory structure and the sizes of the datasets available for main track of the DNS Challenge are:

datasets_fullband 892G
+-- dev_testset 1.7G
+-- impulse_responses 5.9G
+-- noise_fullband 58G
\-- clean_fullband 827G
    +-- emotional_speech 2.4G
    +-- french_speech 62G
    +-- german_speech 319G
    +-- italian_speech 42G
    +-- read_speech 299G
    +-- russian_speech 12G
    +-- spanish_speech 65G
    +-- vctk_wav48_silence_trimmed 27G
    \-- VocalSet_48kHz_mono 974M

In all, you will need about 1TB to store the unpacked data. Archived, the same data takes about 550GB total.

Personalized DNS track

Personalized track shares the noise and IR data with the main track, and the dataset has the following structure:

. 362G
+-- datasets_fullband 64G
|   +-- impulse_responses 5.9G
|   \-- noise_fullband 58G
+-- pdns_training_set 294G
|   +-- enrollment_embeddings 115M
|   +-- enrollment_wav 42G
|   +-- raw/clean 252G
|       +-- english 168G
|       +-- french 2.1G
|       +-- german 53G
|       +-- italian 17G
|       +-- russian 6.8G
|       \-- spanish 5.4G
\-- personalized_dev_testset 3.3G

In all, you will need at least 380GB to store the unpacked data. Archived, the same data takes about 200GB total.

Data checksums

A CSV file containing file sizes and SHA1 checksums for audio clips in both Real-time and Personalized DNS datasets is available at: dns4-datasets-files-sha1.csv.bz2. The archive is 41.3MB in size and can be read in Python like this:

import pandas as pd

sha1sums = pd.read_csv("dns4-datasets-files-sha1.csv.bz2", names=["size", "sha1", "path"])

Code prerequisites

  • Python 3.6 and above
  • Python libraries: soundfile, librosa

NOTE: git LFS is no longer required for DNS Challenge. Please use the download-dns-challenge-4.sh script in this repo to download the data.

Usage:

  1. Install Python libraries
pip3 install soundfile librosa
  1. Clone the repository.
git clone https://github.com/microsoft/DNS-Challenge

  1. Edit noisyspeech_synthesizer.cfg to specify the required parameters described in the file and include the paths to clean speech, noise and impulse response related csv files. Also, specify the paths to the destination directories and store the logs.

  2. Create dataset

python3 noisyspeech_synthesizer_singleprocess.py

Citation:

If you use this dataset in a publication please cite the following paper:

@inproceedings{dubey2022icassp,
  title={ICASSP 2022 Deep Noise Suppression Challenge},
  author={Dubey, Harishchandra and Gopal, Vishak and Cutler, Ross and Matusevych, Sergiy and Braun, Sebastian and Eskimez, Emre Sefik and Thakker, Manthan and Yoshioka, Takuya and Gamper, Hannes and Aichner, Robert},
  booktitle={ICASSP},
  year={2022}
}

The previous challenges were:

@inproceedings{reddy2021interspeech,
  title={INTERSPEECH 2021 Deep Noise Suppression Challenge},
  author={Reddy, Chandan KA and Dubey, Harishchandra and Koishida, Kazuhito and Nair, Arun and Gopal, Vishak and Cutler, Ross and Braun, Sebastian and Gamper, Hannes and Aichner, Robert and Srinivasan, Sriram},
  booktitle={INTERSPEECH},
  year={2021}
}
@inproceedings{reddy2021icassp,
  title={ICASSP 2021 deep noise suppression challenge},
  author={Reddy, Chandan KA and Dubey, Harishchandra and Gopal, Vishak and Cutler, Ross and Braun, Sebastian and Gamper, Hannes and Aichner, Robert and Srinivasan, Sriram},
  booktitle={ICASSP},
  year={2021},
}
@inproceedings{reddy2020interspeech,
  title={The INTERSPEECH 2020 deep noise suppression challenge: Datasets, subjective testing framework, and challenge results},
  author={Reddy, Chandan KA and Gopal, Vishak and Cutler, Ross and Beyrami, Ebrahim and Cheng, Roger and Dubey, Harishchandra and Matusevych, Sergiy and Aichner, Robert and Aazami, Ashkan and Braun, Sebastian and others},
  booktitle={INTERSPEECH},
  year={2020}
}

The baseline NSNet noise suppression:

@inproceedings{9054254,
    author={Y. {Xia} and S. {Braun} and C. K. A. {Reddy} and H. {Dubey} and R. {Cutler} and I. {Tashev}},
    booktitle={ICASSP 2020 - 2020 IEEE International Conference on Acoustics,
    Speech and Signal Processing (ICASSP)},
    title={Weighted Speech Distortion Losses for Neural-Network-Based Real-Time Speech Enhancement},
    year={2020}, volume={}, number={}, pages={871-875},}
@misc{braun2020data,
    title={Data augmentation and loss normalization for deep noise suppression},
    author={Sebastian Braun and Ivan Tashev},
    year={2020},
    eprint={2008.06412},
    archivePrefix={arXiv},
    primaryClass={eess.AS}
}

The P.835 test framework:

@inproceedings{naderi2021crowdsourcing,
  title={Subjective Evaluation of Noise Suppression Algorithms in Crowdsourcing},
  author={Naderi, Babak and Cutler, Ross},
  booktitle={INTERSPEECH},
  year={2021}
}

DNSMOS API: 

@inproceedings{reddy2021dnsmos,
  title={DNSMOS: A Non-Intrusive Perceptual Objective Speech Quality metric to evaluate Noise Suppressors},
  author={Reddy, Chandan KA and Gopal, Vishak and Cutler, Ross},
  booktitle={ICASSP},
  year={2021}
}
@inproceedings{reddy2022dnsmos,
  title={DNSMOS P.835: A non-intrusive perceptual objective speech quality metric to evaluate noise suppressors},
  author={Reddy, Chandan KA and Gopal, Vishak and Cutler, Ross},
  booktitle={ICASSP},
  year={2022}
}

Contributing

This project welcomes contributions and suggestions. Most contributions require you to agree to a Contributor License Agreement (CLA) declaring that you have the right to, and actually do, grant us the rights to use your contribution. For details, visit https://cla.opensource.microsoft.com.

When you submit a pull request, a CLA bot will automatically determine whether you need to provide a CLA and decorate the PR appropriately (e.g., status check, comment). Simply follow the instructions provided by the bot. You will only need to do this once across all repos using our CLA.

This project has adopted the Microsoft Open Source Code of Conduct. For more information see the Code of Conduct FAQ or contact opencode@microsoft.com with any additional questions or comments.

Legal Notices

Microsoft and any contributors grant you a license to the Microsoft documentation and other content in this repository under the Creative Commons Attribution 4.0 International Public License, see the LICENSE file, and grant you a license to any code in the repository under the MIT License, see the LICENSE-CODE file.

Microsoft, Windows, Microsoft Azure and/or other Microsoft products and services referenced in the documentation may be either trademarks or registered trademarks of Microsoft in the United States and/or other countries. The licenses for this project do not grant you rights to use any Microsoft names, logos, or trademarks. Microsoft's general trademark guidelines can be found at http://go.microsoft.com/fwlink/?LinkID=254653.

Privacy information can be found at https://privacy.microsoft.com/en-us/

Microsoft and any contributors reserve all other rights, whether under their respective copyrights, patents, or trademarks, whether by implication, estoppel or otherwise.

Dataset licenses

MICROSOFT PROVIDES THE DATASETS ON AN "AS IS" BASIS. MICROSOFT MAKES NO WARRANTIES, EXPRESS OR IMPLIED, GUARANTEES OR CONDITIONS WITH RESPECT TO YOUR USE OF THE DATASETS. TO THE EXTENT PERMITTED UNDER YOUR LOCAL LAW, MICROSOFT DISCLAIMS ALL LIABILITY FOR ANY DAMAGES OR LOSSES, INLCUDING DIRECT, CONSEQUENTIAL, SPECIAL, INDIRECT, INCIDENTAL OR PUNITIVE, RESULTING FROM YOUR USE OF THE DATASETS.

The datasets are provided under the original terms that Microsoft received such datasets. See below for more information about each dataset.

The datasets used in this project are licensed as follows:

  1. Clean speech:
  1. Noise:
  1. RIR datasets: OpenSLR26 and OpenSLR28:

Code license

MIT License

Copyright (c) Microsoft Corporation.

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE