On the use of phone log-likelihood ratios as features in spoken language recognition

Extended phone log-likelihood ratio features and acoustic-based i-vectors for language recognition

“…In [6] and [7] it is shown that the PLLR features can be successfully used for language and speaker recognition tasks. Its success is probably due to the simplicity of its calculation and because they can be easily integrated with the i-vector framework where the PLLR can be seen as an alternative to the acoustic MFCC-SDC features.…”

“…Its success is probably due to the simplicity of its calculation and because they can be easily integrated with the i-vector framework where the PLLR can be seen as an alternative to the acoustic MFCC-SDC features. On the other hand, as proved in [7], other alternative features as frame-level posteriors or phone log-posteriors (which are usually provided by phone recognizers) are not suitable for tasks where the features are assumed to be Gaussiandistributed. In contrast, the transformation from log posteriors into log-likelihood ratios (LLR) provides final distributions that are nearly Gaussian.…”

“…In contrast, the transformation from log posteriors into log-likelihood ratios (LLR) provides final distributions that are nearly Gaussian. In order to calculate the PLLR features [14], the acoustic posterior probability of a phone unit m at each frame f, is calculated by summing up the posteriors of its corresponding states:…”

“…These phone decoders use a three-state model per phone, which means that three posterior probabilities per unit are given at each frame. Since these posterior probabilities are encoded by default, we applied some simple mathematical formulas to decode them (see section 4.2 in [7] for further details).…”

“…The main change has been the incorporation of a new kind of phonotactic subsystem that uses Phone Log-Likelihood Ratios features (PLLR) that have proved to improve both language [6] and speaker recognition systems [7]. Later, we will show how these features can be extended to provide a better performance thanks to the use of likelihood ratios at a phone state level, instead of a phone level, the addition of new coefficients based on the Shifted Delta Cepstra (SDC) philosophy [8], which we called Shifted Delta PLLR Coefficients (SDPC) and the use of PCA and HLDA dimensionality reduction techniques.…”