Improved Phonetic Speaker Recognition Using Lattice Decoding

Abstract

The current "state-of-the-art" in phonetic speaker recognition uses relative frequencies of phone n-grams as features for training speaker models and for scoring test-target pairs. Typically, these relative frequencies are computed from a simple 1-best phone decoding of the input speech. In this paper, we present results on the Switchboard-2 corpus, where we compare 1-best phone decodings versus lattice phone decodings for the purposes of performing phonetic speaker recognition. The phone decodings are used to compute relative frequencies of phone bigrams, which are then used as inputs for two standard phonetic speaker recognition systems: a system based on log-likelihood ratios (LLRs) [1,2], and a system based on support vector machines (SVMs) [3]. In each experiment, the lattice phone decodings achieve relative reductions in equal-error rate (EER) of between 31% and 66% below the EERs of the 1-best phone decodings. Our best phonetic system achieves an EER of 2.0% on 8-conversation training and 1.4% when combined with a GMM-based system.

Statistical lattice-based spoken document retrieval

“…Our use of expected counts differs from Saraclar and Sproat [2004] in that we estimate probability models from the expected counts. Conceptually, our method of estimating language models from expected term frequencies is close to that of Hatch et al [2005] and that of . In practice, however, our method differs from both works in a number of ways.…”

“…In practice, however, our method differs from both works in a number of ways. While Hatch et al [2005] derive phone bigram statistics for representing phonotactics, we derive word statistics for representing semantics. In addition, while estimate the probability of unseen phone ngrams using lower order n-gram statistics, this smoothing approach is inapplicable in our case, as we derive word models where the model vocabulary is large and where the sparse data problem is of a different nature.…”

“…Expected counts have also been used to summarize the phonotactics of a speech recording represented in a lattice: Hatch et al [2005] performed speaker recognition by computing the expected counts of phone bigrams in a phone lattice, and estimating an unsmoothed probability distribution of phone bigrams. In addition, Allauzen et al [2003Allauzen et al [ , 2004a outlined algorithms based on automata theory for obtaining expected counts and estimating backoff and interpolated language models from such counts, for speech recognizer adaptation and speech mining.…”

Nonparametric feature normalization for SVM-based speaker verification

“…The strongest motivation for it comes from two related properties: any feature distribution, to the extent that it matches the background distribution, is warped to a uniform distribution over the interval 0; 1]. Conversely, the kernel-induced distance between datapoints D(x; y) 2 = K(x; x) + K(y; y) 2K(x;y) = jjx yjj 2 (6) (in the case of a linear kernel K(x; y)) is such that along any single feature dimension, two points x and y are separated by a distance proportional to the number of background data samples falling between x and y. In other words, the normalization stretches the feature space in areas of high population density and shrinks it in areas of low density.…”

“…Here we use the phone recognition-based modeling paradigm of [3] with the lattice-based refinement of [6]. An English open-loop phone recognizer is run on each conversation side, generating lattices.…”