Time and Frequency Pruning for Speaker Identification - Archive ouverte HAL Accéder directement au contenu
Communication Dans Un Congrès Année : 1998

Time and Frequency Pruning for Speaker Identification

L Besacier
  • Fonction : Auteur
  • PersonId : 1048997
J.-F. Bonastre

Résumé

This work is an attempt to refine decisions in speaker identification. A test utterance is divided into multiple time-frequency blocks on which a normalized likelihood score is calculated. Instead of averaging the block-likelihoods along the whole test utterance, some of them are rejected (pruning) and the final score is computed with a limited number of time-frequency blocks. The results obtained in the special case of time pruning lead the authors to experiment a joint time and frequency pruning approach. The optimal percentage of blocks pruned is learned on a tuning data set with the minimum identification error criterion. Validation of the time-frequency pruning process on 567 speakers leads to a significant error rate reduction (up to 41% reduction on TIMIT) for short training and test duration. ,QWURGXFWLRQ Mono-gaussian models for speaker recognition have been largely replaced by Gaussian Mixture Models (GMM) which are dedicated to modeling smaller clusters of speech. The Gaussian mixture modeling can be seen as a FRRSHUDWLRQ of models since the gaussian mixture density is a weighted linear combination of uni-modal gaussian densities. The work presented here is rather concerned with FRPSHWLWLRQ of models since different mono-gaussian models (corresponding to different frequency subbands) are applied to the test signal and the decision is made with the best or the N-best model scores. More precisely, a test utterance is divided into time-frequency blocks, each of them corresponding to a particular frequency subband and a particular time segment. During the recognition phase, the block scores are accumulated over the whole test utterance to compute a global score and take a final decision. In this work, we investigate accumulation using a hard threshold approach since some block scores are eliminated (pruning) and the final decision is taken with a subset of these scores. This approach should be robust in the case of a time-frequency localized noise since the least reliable time-frequency blocks can be removed. Even in the case of clean speech, some speaker test utterance blocks can be simply more similar to another speaker model than to the target speaker model itself. Removing these error-prone blocks should lead to a more robust decision. In 6HFWLRQ , a formalism is proposed to describe our block-based speaker recognition system. The potential of this approach is shown with a special case of the formalism: time pruning (6HFWLRQ). Experiments intended to find the optimal percentage of blocks pruned are described in 6HFWLRQ. The optimal parameters (percentage of blocks pruned) are validated on TIMIT and NTIMIT databases (6HFWLRQ). Finally, we summarize our main results and outline the potential advantages of the time-frequency pruning procedure in 6HFWLRQ .)RUPDOLVP 0RQRJDXVVLDQ µVHJPHQWDO ¶ PRGHOLQJ Let { } [ W W 0 1≤ ≤ be a sequence of M vectors resulting from the S-dimensional acoustic analysis of a speech signal uttered by speaker X. These vectors are summarized by the mean vector [ and the covariance matrix X: [ 0 [ ; 0 [ [ [ [
Fichier principal
Vignette du fichier
icpr98.pdf (81.27 Ko) Télécharger le fichier
Origine : Fichiers produits par l'(les) auteur(s)
Loading...

Dates et versions

hal-02157124 , version 1 (15-06-2019)

Identifiants

  • HAL Id : hal-02157124 , version 1

Citer

L Besacier, J.-F. Bonastre. Time and Frequency Pruning for Speaker Identification. International Conference on Pattern Recognition, 1998, brisbane, Australia. ⟨hal-02157124⟩

Collections

UNIV-AVIGNON LIA
33 Consultations
59 Téléchargements

Partager

Gmail Facebook X LinkedIn More