Key challenges and results

The main strength of OCTAVE is the use of voice biometrics in a fashion that provides high recognition accuracy whilst offering reliability in unsupervised access control applications, whether onsite or remote.

To reach this objective, a number of key challenges must be met.

To improve the level of confidence towards the deployment of the current voice biometric methods in real business environments

OCTAVE is delivering an access control platform through a trusted biometric authentication service (TBAS) to use in data-sensitive and mission-critical applications in real business/commercial environments. Two main contexts have been detailed through uses cases and functional, security, performance, availability and scalability requirements:
1. authentication of logical access to remote banking services at Findomestic (highly data-sensitive)
2. authentication of physical access to restricted sensitive areas at SEA-Aeroporti di Milano (highly mission critical)
While simultaneously relieving end-users from the inconvenience of dealing with textual passwords, the OCTAVE platform reduces the economic and practical burden related to password loss and recovery: through the use of biometrics, OCTAVE is delivering operational, interoperable technology, installed and tested in a broad-band network, managing secure access to diverse and trustworthy smart services whereby privacy is guaranteed by design.

Protection of personal data

The use of biometrics raises certain privacy concerns: persons or organizations that collect and manage personal information must protect it from misuse and must respect certain rights of the data owners which are guaranteed by EU law. Technical and legal aspects of the safety, security and personal data protection, with respect to the development of the TBAS platform and putting a special emphasis on digital biometric data, have been surveyed and addressed, in order to properly approach the implementation of the OCTAVE solution.

Provision of a user authentication platform that be capable of communicating with applications in a standard way

Authentication must be provided to, and only to, the legitimate requesting user: starting from the state-of-the-art systems and enriching it with a study of ASV system exposure to spoofing, this work validated the need to strengthen ASV systems with spoofing countermeasures – all systems were shown to be vulnerable – and helped to prioritise the specific weaknesses to be addressed with priority.
When the nature of spoofing attack is known “a priori”, then other attacks similar in nature can be detected with ease. On the other side, when the nature of the attack varies, countermeasure performance can deteriorate catastrophically.
Since the nature of the attack will never be known “a priori”, either in any of the OCTAVE use case scenarios, or indeed in any foreseeable future application, robustness to unforeseen attacks is paramount and is being addressed.

Matching speaker verification requirements to different typical applications

The optimal Convenience-Security trade-off is reached by means of an hybrid approaches to ASV that combines and fuses different verification algorithms.  The use of single mode text-dependent, text-prompted and text-independent ASV and then fused-mode, i.e. hybrid ASV systems, leads to an acceptable recognition performance without compromising the user convenience.

Language independence

In order to facilitate deployment throughout Europe, the challenge is to deliver technology that performs well across different languages. Therefore a survey of linguistic resources available and useful to OCTAVE objectives has been accomplished, dealing with a total of about 30 datasets: the Italian ones (SIVA, CALI, CLIPS, APASCI, PRIMULA, etc.); the English ones (TIMIT, YOHO, RSR2015, RedDots, etc.); and the multi-language ones (e.g. POLICOST250).
Moreover, the necessary standards and formats required for the “corpus of corpora” have been defined, in order to collect several speech corpora in a unique “normalised dataset”

Speaker verification performance under different communication
channel behaviour

Speaker verification has the advantage of being operational over a wide range of telecommunication channels but these channels, in the real-world, have to deal with environmental adverse conditions like presence of acoustic noises and channel degradation.
Noise disturbances management has been addressed and tested leveraging on a dataset of noise complying with the typical environments of OCTAVE scenarios, e.g.:
1. truck noise
2. airplane noise
On top of this, innovative audio acquisition devices to cope with the unresolved problem of environmental robustness by means of dual-microphone (throat and acoustic) recordings have been experienced.