Conditional pronunciation modeling in speaker detection

Abstract

In this paper, we present a conditional pronunciation niodeling method for the speaker detection task that does not rely on acoustic vectors. Aiming at exploiting higherlevel information carried by the speech signal, it uses timealigned streams of phones and phonemes to model a speaker's specific Pronunciation. Our system uses phonemes drawn from a lexicon of pronunciations of words recognized by an automatic speech recognition system to generate the phoneme stream and an open-loop phone recognizer to generate a phone stream. The phoneme and phone streams are aligned at the frame level and conditional probabilities of a phone, given a phoneme, are estimated using c*occnrrence counts. A likelihood detector is then applied to these probabilities.Performance is measured using the NIST Extended Data paradigm and the Switchboard-I corpus. Using 8 training conversations for enrollment, a 2.1% equal error rate was achieved. Extensions and alternatives, as well as fusion experiments, are presented and discussed.

A new adaptation approach to high-level speaker-model creation in speaker verification

“…The results show that there is significant benefit of fusing high-and low-level features for speaker verification. Among the high-level features investigated, the conditional pronunciation modeling (CPM) technique [30] that extracts multilingual phone sequences from utterances achieves the best performance [20]. One limitation of the CPM in [30] is that it requires multi-lingual corpora to build speaker and background models.…”

“…Among the high-level features investigated, the conditional pronunciation modeling (CPM) technique [30] that extracts multilingual phone sequences from utterances achieves the best performance [20]. One limitation of the CPM in [30] is that it requires multi-lingual corpora to build speaker and background models. To overcome this limitation, Leung et al [32] proposed using articulatory feature (AF) streams to construct CPM and called the resulting models AFCPM.…”

“…These works have led to extensive investigations into high-level features, in which prosodic features [18,[21][22][23][24][25], pronunciation features [26][27][28][29][30], and idiolect features [19,31] were proposed and combined with acoustic features. The results show that there is significant benefit of fusing high-and low-level features for speaker verification.…”

Articulatory-feature based sequence kernel for high-level speaker verification

“…Textindependent speaker verification systems typically extract speaker features from short-term spectra of speech signals to build speaker-dependent Gaussian mixture models (GMMs) [1]. Studies have shown that combining low-level acoustic information with high-level speaker information-such as the usage or duration of particular words, prosodic features and articulatory features (AF)-can improve speaker verification performance [2][3][4][5][6].…”

Improved Phonetic Speaker Recognition Using Lattice Decoding

“…This line of research, which is generally referred to as phonetic speaker recognition, was pioneered by Andrews et al, who used relative frequencies of phone n-grams to capture sequential patterns in an individual's speech [1,2]. This work was subsequently extended in various papers, such as the work of the "SuperSID" team at the JHU 2002 Summer Workshop [5,6,7]. In 2003, Campbell et al used support vector machines (SVMs) to train phonetic speaker models [3].…”