This document accompanies a related collection of audio material referred to as the OCTAVE spoofing corpus. In detail, the spoofing corpus reported here contains a representative sample of replayed audio material originating from a number of varied playback and re-capture devices. Besides the audio material, the corpus definition contains a set of pre-configured evaluation protocols containing partitioning of the data into training and evaluation parts and consisting of two related, but disjoint, tasks: (a) speaker verification under replay spoofing, and (b) standalone spoofing detection. This accompanying documentation explains, firstly, the rationale of focusing on replay spoofing as opposed to other, more advanced spoofing attacks such as speech synthesis or voice conversion. It then describes how the specific replay data was defined, collected and post-processed to define the audio files and how this data was partitioned to define the protocols. In short, the data collection consisted of a voluntary collection effort by multiple “friendly attackers” recruited from within the OCTAVE project members to provide as wide array of varied replay devices, recapture devices and environments as possible. This particular, crowd-sourced effort to collect variable data represents, in fact, a unique feature of the collected spoofing corpus in relation to earlier off-the-shelf replay corpora available prior to OCTAVE that have usually contained relatively homogenous conditions. The deliverable provides a tentative validation of the collected data by reporting preliminary speaker verification and stand-alone spoofing countermeasure results.
Source: WP 7 Test and Verification
Dissemination level: Confidential
To know more about the document, you may place a request in the ‘Contact’ section of this site. We reserve the right to decide how much we can disclose.