Analysis of MLP-Based Hierarchical Phoneme Posterior Probability Estimator

Abstract

Abstract-We analyze a simple hierarchical architecture consisting of two multilayer perceptron (MLP) classifiers in tandem to estimate the phonetic class conditional probabilities. In this hierarchical setup, the first MLP classifier is trained using standard acoustic features. The second MLP is trained using the posterior probabilities of phonemes estimated by the first, but with a long temporal context of around 150-230 ms. Through extensive phoneme recognition experiments, and the analysis of the trained second MLP using Volterra series, we show that (a) the hierarchical system yields higher phoneme recognition accuracies -an absolute improvement of 3.5% and 9.3% on TIMIT and CTS respectively -over the conventional single MLP based system, (b) there exists useful information in the temporal trajectories of the posterior feature space, spanning around 230 ms of context, (c) the second MLP learns the phonetic temporal patterns in the posterior features, which include the phonetic confusions at the output of the first MLP as well as the phonotactics of the language as observed in the training data, and (d) the second MLP classifier requires fewer number of parameters and can be trained using lesser amount of training data.

Phone recognition experiments with 2D-DCT spectro-temporal features

“…If this second stage net is trained on several neighboring frames -similar to the MFCC-based net -then it is able to correct some of the errors of the lower stage net(s) with the help of the long-term context. Hence, applying such a second stage network is already useful in itself, as was recently shown in [15] or [16]. We will compare our earlier results with two 2-stage configurations: the first is trained only on the MFCC-based posteriors, while the second combines the MFCC-based and the 2D-DCT based probabilities.…”

“…Both physiological and psychoacoustic experimental results indicate that the human brain extracts information from much longer time spans. Technically the simplest solution for this is to work with larger windows along the timeaxis: in neural-net based recognizers it is now standard practice to train the system on 9 or more neighboring MFCC vectors [7,15,16]. However, there is also evidence that the brain processes relatively narrow frequency bands quasi-separately [2,6].…”

Integrating articulatory features using Kullback-Leibler divergence based acoustic model for phoneme recognition

“…The 61 hand labeled phonetic symbols are mapped to set of 39 phonemes with an additional garbage class [7]. The experimental setup is exactly same as the one described in [8]. All the MLPs (for phoneme posterior and articulatory posterior estimation) use the PLP cepstral coefficients with a context window of 9 frames as input.…”

“…The size of the hidden layer of all the MLPs is determined by fixing the total number of parameters to 35% of the training data following the previous work [8]. The articulatory posteriors and phoneme posteriors are estimated from MLP trained using ICSI Quicknet software 1 .…”

“…This approach showed improvements in articulatory feature classification compared to an equivalent system where they were treated independently. Motivated from the previous studies [4], [10], and the hierarchical MLP framework [8], we investigate a novel multi-stage MLP classifier based approach to model the inter-feature dependencies of articulatory features.…”