J. A. Bilmes and K. Kirchhoff. Directed graphical models of classifier combinations: application to phone recognition. In ICSLP, Beijing, China, October 2000. Home/Search Document Details and Download
Summary Related Articles Check |
This paper is cited in the following contexts:
Multi-Stream ASR Trained with Heterogeneous Reverberant Environments - Shire (Correct)
....and T60. Reverberated data was applied artificially to the speech via convolution with the room impulse response. 3. SIMPLE COMBINATION RESULTS There are a number of ways to combine probability estimates from several classifiers. These are usually based upon different independence assumptions [1]. A common method used by researchers is to average the logarithm of the probabilities, P #qjx1;x 2 ; x N#= exp# P i=1 w i log P #qjx i ## P q 0 exp# P i=1 w i log P #q 0 jx i ## where P #qjx i # is the posterior probability of class q given features x of stream i, and the weights w i ....
J. A. Bilmes and K. Kirchhoff. Directed graphical models of classifier combinations: application to phone recognition. In ICSLP, Beijing, China, October 2000.
Combination And Joint Training Of Acoustic Classifiers For.. - Kirchhoff, Bilmes (2000) (2 citations) Self-citation (Bilmes Kirchhoff) (Correct)
....(the mean rule) computes a weighted average of classifier outputs. Another method (the product rule) multiplies and then renormalizes these probabilities. Other techniques compute the maximum, minimum, or median of the classifier outputs [22] Other methodologies combine using statistical models [5] or jointly train separate classifiers [28, 14] In ASR, classifier combination can occur at different levels including the feature stream [3, 19, 12, 20] the HMM state [21] or at higher levels such as at the syllable [32] or sentence [11] In ASR, HMM state level combination (e.g. combining ....
.... linear combinations of the output distributions (e.g. by averaging over, or multiplying, the outputs) 13, 22] combining the outputs by a higher level classifier, Dempster Shafer theory [29] and majority voting [24] An alternative approach is to model a dependence between classification errors [6, 31, 5]. Many studies on classifier combination focus on what we call single level classification. A typical single level classification system consists of two components: a featureextraction component which maps the input signal to feature vectors, and a classification component which assigns a class ....
[Article contains additional citation context not shown here]
J.A. Bilmes and K. Kirchhoff. Directed graphical models of classifier combination: Application to phone recognition. In Proc. Int. Conf. on Spoken Language Processing, Beijing, China, 2000.
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