Vowel-category based Short Utterance Speaker Recognition

Abstract

The impact of Short Utterances in Speaker Recognition is of significant importance. Despite the advancements in short utterance speaker recognition (SUSR), text dependence and the role of phonemes in carrying speaker information needs further investigation. This paper presents a novel method of using vowel categories for SUSR. We define Vowel Categories (VC's) considering Chinese and English languages. After recognition and extraction of phonemes, the obtained vowels are divided into VC's, which are then used to develop Universal Background VC Models (UBVCM) for each VC. Conventional GMM-UBM system is used for training and testing. The proposed categories give minimum EERs of 13.76%, 14.03% and 16.18% for 3, 2 and 1 second respectively. Experimental results show that in text dependent SUSR, significant speaker-specific information is present at phoneme level. The similar properties of phonemes can be used such that accurate speech recognition is not required, rather Phoneme Categories can be used effectively for SUSR. Also, it is shown that vowels contain large amount of speaker information, which remains undisturbed when VC are employed.

Speaker verification with short utterances: a review of challenges, trends and opportunities

“…A systematic combination of SUVN with LDA and source-normalised (SN)-LDA was further used effectively. Moreover, an alternative approach was introduced using PLDA to directly model the SUV which showed that for the combination of [44] performance evaluation JFA, i-vector CSS, WCCN, LDA, NAP, SDNAP, GPLDA [7] calibration evaluation linear calibration, cosine kernel, normalised cosine kernel 2012 [80] performance evaluation inclusion of short utterances in development data-set [51] performance evaluation an ad hoc fusion system of different TV spaces [81] evaluation of phoneme effects adding phonetic information, WCCN and EFR [82] adding of phonetic information VCs, UBVCM [83] adding of syllable information syllable categories, universal background syllable models 2013 [49] analysis on phoneme distribution score calibration with log duration as QMF, synthetic i-vectors [84] analysis of phonetic content TD-ASV, multiple enrolment, used speaker and phonetic content [85] analysis on confusion errors finding speaker-specific phonemes, formulate text using unique phonemes [21] analysis on score calibration QMFs, stacked scores, shared scaling, extrapolation [86] performance analysis TV, PLDA [87] source and utterance -dur. norm.…”

“…The paper [82] introduced an approach of using vowel categories (VCs). After recognising and extracting the phonemes, extracted vowels are divided into VC's to generate universal background VC models (UBVCMs) for each VC.…”

Multi-scale kernels for short utterance speaker recognition

“…Kanagasundaram et al proposed a source and utteranceduration normalized linear discriminant analysis approach to compensate session variability in short utterance i-vector systems [2]. Meanwhile, there are alternate approaches which resort to the text-dependant information, such as vowel-category information [3,4], multi-layer acoustic and temporal structure information [5,6], and so on.…”