Redes neurais profundas com fusão de características na verificação biométrica pela voz
Nenhuma Miniatura disponível
Data
2022-04-14
Autores
Título da Revista
ISSN da Revista
Título de Volume
Editor
Universidade Federal de Goiás
Resumo
The voice spoken by a person has a considerable variability which is due both to factors
related to the speaker himself, such as accent, emotional state, and voice transition over age,
as well as on factors external to the speaker, such as background noise, reverberation,
capture equipment, and the digitalization process. Therefore, there are many challenges
present in the task of biometric verification by voice. The use of neural networks to tackle this
problem brought a big leap in performance when compared to previous techniques, and the
main input format used is the spectrogram. For voices, the spectrogram can emphasize
different characteristics depending on the generation parameters. The purpose of this work is
to explore feature fusion in biometric verification by voice, particularly with by using dual
spectrograms as input to the model. This approach is justified by the existence of works that
also use it in other tasks related to voice and speech, such as keyword spotting, detection of
voiced excerpts and musical classification. From the results, it was possible to validate the
hypothesis that the use of dual spectrograms allows a performance gain in existing models,
implying that certain types of spectrogram carry complementary information. The Equal Error
Rate obtained was 1.61 for the model trained with dual spectrograms, which is 26% less than
the EER rate of 2.22 obtained by the reference work [Chung et al. 2020]. Furthermore, the
model proposed in this work has better performance for any decision threshold when
compared to the reference work, either to minimize false positives or false negatives.
Descrição
Palavras-chave
Citação
VIRGILLI, R. Redes neurais profundas com fusão de características na verificação biométrica pela voz. 2022. 63 f. Dissertação (Metrado em Ciência da Computação) - Universidade Federal de Goiás, Goiânia, 2022.