MULTISPEECH team member is co-organizing the third edition of spoofed/fake audio detection challenge. The task is to design automatic systems capable of discriminating between natural and spoofed speech. The official announcement of the challenge can be found below.
*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*
ASVspoof 2019 CHALLENGE:
Future horizons in spoofed/fake audio detection
http://www.asvspoof.org/
=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=
Can you distinguish computer-generated or replayed speech from authentic/bona fide speech? Are you able to design algorithms to detect spoofs/fakes automatically?
Are you concerned with the security of voice-driven interfaces?
Are you searching for new challenges in machine learning and signal processing?
Join ASVspoof 2019 – the effort to develop next-generation countermeasures for the automatic detection of spoofed/fake audio. In combining the forces of leading research institutes and industry, ASVspoof 2019 encompasses two separate sub-challenges in logical and physical access control, and provides a common database of the most advanced spoofing attacks to date. The aim is to study both the limits and opportunities of spoofing countermeasures in the context of automatic speaker verification and fake audio detection.
CHALLENGE TASK:
Given a short audio clip, determine whether it represents authentic/bona fide human speech, or a spoof/fake (replay, synthesized speech or converted voice). You will be provided with a large database of labelled training and development data and will develop machine learning and signal processing countermeasures to distinguish automatically between the two. Countermeasure performance will be evaluated jointly with an automatic speaker verification (ASV) system provided by the organisers.
BACKGROUND:
The ASVspoof 2019 challenge follows on from two previous ASVspoof challenges, held in 2015 and 2017. The 2015 edition focused on spoofed speech generated with text-to-speech (TTS) and voice conversion (VC) technologies. The 2017 edition focused on replay spoofing. The 2019 edition is the first to address all three forms attack and the latest, cutting-edge spoofing attack technology.
ADVANCES:
Today’s state-of-the-art, TTS and VC technologies produce speech signals that are as good as perceptually indistinguishable from bona fide speech. The LOGICAL ACCESS sub-challenge aims to determine whether the advances in TTS and VC pose a greater threat to the reliability of automatic speaker verification and spoofing countermeasure technologies. The PHYSICAL ACCESS sub-challenge builds upon the 2017 edition with a far more controlled evaluation setup which extends the focus of ASVspoof to fake audio detection in, e.g. the manipulation of voice-driven interfaces (smart speakers).
METRICS:
The 2019 edition also adopts a new metric, the tandem detection cost function (t-DCF). Adoption of the t-DCF metric aligns ASVspoof more closely to the field of ASV. The challenge nonetheless focuses on the development of standalone spoofing countermeasures; participation in ASVspoof 2019 does NOT require any expertise in ASV. The equal error rate (EER) used in previous editions remains as a secondary metric, supporting the wider implications of ASVspoof involving fake audio detection.
SCHEDULE:
Training and development data release: 19th December 2018
Participant registration deadline: 8th February 2019
Evaluation data release: 15th February 2019
Deadline to submit evaluation scores: 22nd February 2019
Organisers return results to participants: 15th March 2019
INTERSPEECH paper submission deadline: 29th March 2019
REGISTRATION:
Registration should be performed once only for each participating entity and by sending an email to registration@asvspoof.org with ‘ASVspoof 2019 registration’ as the subject line. The mail body should include: (i) the name of the team; (ii) the name of the contact person; (iii) their country; (iv) their status (academic/non-academic), and (v) the challenge scenario(s) for which they wish to participate (indicative only). Data download links will be communicated to registered contact persons only.
MAILING LIST:
Subscribe to general mailing list by sending e-mail with subject line ‘subscribe asvspoof2019’ to sympa@asvspoof.org
To post messages to the mailing list itself, send e-mails to asvspoof2019@asvspoof.org
ORGANIZERS*:
Junichi Yamagishi, NII, Japan & Univ. of Edinburgh, UK
Massimiliano Todisco, EURECOM, France
Md Sahidullah, Inria, France
Héctor Delgado, EURECOM, France
Xin Wang, National Institute of Informatics, Japan
Nicholas Evans, EURECOM, France
Tomi Kinnunen, University of Eastern Finland, Finland
Kong Aik Lee, NEC, JAPAN
Ville Vestman, University of Eastern Finland, Finland
(*)Equal contribution
CONTRIBUTORS:
University of Edinburgh, UK
Nagoya University, Japan
University of Science and Technology of China, China
iFlytek Research, China
Saarland University / DFKI GmbH, Germany
Trinity College Dublin, Ireland
NTT Communication Science Laboratories, Japan
HOYA, Japan
Google LLC (Text-to-Speech team, Google Brain team, Deepmind)
University of Avignon, France
Aalto University, Finland
University of Eastern Finland, Finland
EURECOM, France
FURTHER INFORMATION:
info@asvspoof.org
