Burshtein, D. Robust Parametric Modeling of Durations in Hidden Markov Models. In Proceedings of ICASSP '95 (Detroit, MI, May 1995), pp. 548-551.  Home/Search   Document Not in Database   Summary   Related Articles   Check

.... h [ hc S ( S ( argmax cVr S ( with cVr S ( pVrS( c pc( pVrS( pVrS( pVrS( c pc( pc( The phone duration criterion The probability to observe a phoneme of duration has been modelized by a gamma density function [5] for each french phoneme, using a training corpus of about 8500 hand labelled segments. The obtained modelization is not context dependent and does not take the phone position in the rhythmic group into consideration. However, an automatic duration normalization is currently under development. ....

....The corresponding value of the threshold is 0.99. This demonstrates the reliability of the prediction algorithm and proves that the global duration constraint is very efficient. As an example, for the voiced region of Fig. 1, the global duration constraint predicts the confidence interval [5,12], which allows the system to focus on only 1491 paths among the initial 389 590 paths. Though the risk (1 ) is minimal, the case where no path satisfies this constraint is still possible. To avoid this irremediable error, we have implemented a retrieving technique which, in this case, selects ....

Burshtein D. (1995) "Robust parametric modeling of durations in hidden Markov models", Proc. ICASSP-95, pp. 548-551

....distribution of state dimensions to have a maximum at around 10 pixels for the horizontal dimension of contiguous red pixels, and around 20 pixels for the vertical dimension. For modeling durations, the Gamma distribution was found to produce a good fit to observed duration histograms [3]. It assigns zero probabilities to negative values of the variable, which is very desirable for modeling durations or dimensions. This distribution is of 6 the form p(x) ff p Gamma(p) e Gammaffx x p Gamma1 (11) The two free parameters of the distribution are ff and p. For this ....

Burshtein, D., "Robust Parametric Modeling of Durations in Hidden Markov Models", IEEE, 1995.

....the words can not be predicted. Typical state duration models yield a Geometric distribution of state occupation, with rapidly decreasing likelihood of remaining in a given state as time increases. This shortcoming has been addressed before, by applying duration models based on Gamma distributions [15] as well as duration probabilities based directly on the training data [16] In our implementation, transition probabilities were set to be all equally likely, so that no assumptions were made about the a priori likelihood of one category following another category. In order to make use of a ....

Burshtein, D., "Robust Parametric Modeling of Durations in Hidden Markov Models", ICASSP-95, Detroit, pp. 548-551, 1995.

....speech rate. 2. 2 CSR Studies Pallett et al. 1994) showed clearly that the presence of fast speech causes a degradation in performance of automatic recognition [25] Most researchers that have attempted to compensate for fast speech use phone duration modeling either during recognition [34] 21][3] or in post processing [22] 2] Their methods do not appear to improve recognition significantly. However, they all agree that the state transition probabilities inherent in acoustic modeling using hidden markov models are problematic with fast speech. In addition, Anastasakos, et al. 1995) ....

....stops will have a much higher phone rate than a sentence containing many stressed vowels and fricatives. Figure 4 3 Histograms for the duration of two different phones. One common alternative is to base the rate of speech on comparisons of observed phone duration with predicted phone duration [2][3][21] 34] The duration of each phone (or phone in context) can then be modelled with a probability density function (PDF) estimated from the training data. The Log Normal [21] Gamma [7] 3] and Poisson distributions have all been used in duration modelling, but it is important to note that the ....

[Article contains additional citation context not shown here]

D. Burshtein [1995], "Robust Parametric Modeling of Durations in Hidden Markov Models," ICASSP, Vol. 1, pp. 548-551.

....data. Table 1 summarizes the architecture and some training parameters for the two types of phone networks. 3.2. Phone Recognition Phone recognition is performed on the OGI TS test set using Viterbi search. CI decoding uses both a bigram language model and gamma duration models as described in [4]. CD decoding uses only a bigram language model plus the generalized biphones constraint. The various thresholds and weightings of the language model and duration models are determined empirically using the validation data. The insertions are kept to 10 by adjusting the transition penalty. 1 ....

D. Burshtein. "Robust Parametric Modeling of Durations in Hidden Markov Models". Proceedings of IEEE ICASSP, pages 548--551, May 1995.

No context found.

Burshtein, D. Robust Parametric Modeling of Durations in Hidden Markov Models. In Proceedings of ICASSP '95 (Detroit, MI, May 1995), pp. 548-551.

No context found.

D. Burshtein, "Robust parametric modeling of durations in hidden Markov Models," IEEE Trans. Speech Audio Processing, vol. 4, pp. 240--242, May 1996.

No context found.

D. Burshtein. Robust parametric modeling of durations in hidden Markov models. In Proc. of ICASSP, volume 1, pages 548--551, Detroit, Michigan U.S.A, May 1995.

No context found.

D. Burshtein. Robust parametric modeling of durations in hidden Markov models. In Proc. of ICASSP, volume 1, pages 548--551, Detroit, Michigan U.S.A, May 1995.

No context found.

D. Burshtein. Robust parametric modeling of durations in hidden markov models. In Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 1995.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC