This deliverable represents optimization and evaluation of single mode voice biometric engines defined in the earlier OCTAVE project deliverable D11. To this end, we extend the speaker verification evaluation protocols and then execute extended experiments to enhance the robustness and operability of the earlier baseline system. Specifically, we improve both feature extraction (introduction of new features) and back-end speaker modelling components (addition of i-vector and HMM recognizer, optimization of acoustic model parameters, addition of score normalization) to enhance speaker verification performance. Further, with the motivations to verify verbal content besides the speaker identity, to prevent replay attacks and to ensure high quality of enrolment utterances, several new methods for utterance verification (UV) are implemented and compared on specifically designed evaluation protocol. In addition, noting that most state-of-the-art speaker verification systems assume knowledge of the speaker’s gender either at enrolment or verification stage or require training of gender-specific universal background models, we address the question of gender-independent speaker verification where such requirements are not necessary; this includes implementation of new gender detection modules. Finally, studies on the effects of signal bandwidth, duration and language mismatch between the development and operational languages are addressed.
Source: WP 4 Hybrid Voice Biometrics
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.