MSA

A Benchmark for Multi-speaker Anonymization

This is an implementation of the paper - A Benchmark for Multi-speaker Anonymization

The authors are Xiaoxiao Miao, Ruijie Tao, Chang Zeng, Xin Wang.

Audio samples can be found here: https://xiaoxiaomiao323.github.io/msa-audio/

Dependencies

git clone https://github.com/xiaoxiaomiao323/MSA.git

cd MSA

bash install.sh

bash demo.sh

Demo Results

Three randomly selected conversations with 3, 4, and 5 speakers respectively are anonymized using different MSAs.

	DER	FA	MS	SC
Original	5.54	0.00	0.00	5.54
Resynthesized	7.28	0.20	0.17	6.91
$A_{selec}$	6.05	0.00	0.00	6.05
$A_{ds}$	6.49	0.43	0.00	6.06
$A_{as}$	6.06	0.00	0.00	6.06

The results of $A_{selec}$, $A_{ds}$, and $A_{as}$ are similar using only three conversations. If you try more samples with more speakers, $A_{as}$ will show better DERs at a statistical level.

Instructions for aonymization for your own conversations

Put your data under "data/wavs"
[Optional] If you have real RTTMs, put it under "data/rttms" (only needed if you want to compute DERs)
bash demo.sh

Acknowledgments

This study is partially supported by JST, PRESTO Grant Number JPMJPR23P9, Japan and SIT-ICT Academic Discretionary Fund.

License

The anon/adapted_from_facebookreaserch subfolder has Attribution-NonCommercial 4.0 International License. The anon/adapted_from_speechbrain subfolder has Apache License. They were created by the facebookreasearch and speechbrain orgnization, respectively. The anon/scripts and anon /anon_control subfolder has the MIT license.

Because this source code was adapted from the facebookresearch and speechbrain, the whole project follows
the Attribution-NonCommercial 4.0 International License.