Variability compensation for speaker verification with short utterances

Flavio Jorge Reyes Diaz, Gabriel Hernández-Sierra, José R. Calvo de Lara

Resumen


Nowadays, represents an attractive challenge the application of Automatic speaker recognition in real scenarios, where the use of short duration signals for forensic or biometric speaker verification is very common. In this paper we perform an analysis of the behavior of within-class and between-classes scatter matrices, showing the importance to reduce within-class scatter to face the speaker recognition with short duration utterances. In addition, two duration compensation methods for short duration utterances on i-vector framework were proposed. Both of them were evaluated through speaker verification experiments on NIST-SRE 2008 dataset. The proposed methods shown an improvements under enrollment-test matched conditions regard to the duration.

Palabras clave


short utterance; variability compensation; speaker verification; i-vector

Texto completo:

PDF (English)

Enlaces refback

  • No hay ningún enlace refback.




_________________________________________________________________________________________________________

La Universidad de las Ciencias Informáticas (UCI), a través del sello editorial Ediciones Futuro, publica los contenidos de la Revista Cubana de Ciencias Informáticas (RCCI) bajo licencia Creative Commons de tipo Atribución 4.0 Internacional (CC BY 4.0). Esta licencia permite a otros distribuir, mezclar, ajustar y construir a partir de su obra, incluso con fines comerciales, siempre que le sea reconocida la autoría de la creación original.
_________________________________________________________________________________________________________

 INDEXACIÓN