H. Bourland and N. Morgan. Connectionist Speech Recognition-A Hybrid Approach. Kluwer Academic, 1994.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....the main parts of the implementation example of figure 5, a few definitions are needed. The SpatioTemporal Neural Network (STNN) blocks are implemented by Multi Layer Perceptrons (MLPs) with a tapped delay line as input, a sigmoid activation function and the relative entropy as cost function [12]. The Self Modularization Neural Network (SMNN) block of figure 4 is composed of a STNN block followed by the decimation operator. Each SMNN represents a specific tier of the MNN architecture example. Decimation operator STNN h (n) a D Figure 4. SMNN building block. 4.2 Implementation ....

....decoding has been assembled to create these phonemes, sub words and words target vectors. Training Neural Networks on speech material is somehow a particular case, the large size of the specific databases and their redundant content has led researchers to adopt domain specific training procedures [12]. A common example is summarized as follows. The MLP training is based on speech frames. Each window of a fixed number of speech frames is associated with a target vector. Their performance is measured on a PCF basis, Percentage of the Correct Frames from the total number of frames. Only a ....

Herv e Bourlard, Nelson Morgan. Connectionist Speech Recognition: a Hybrid Approach. Kluwer Academic Publishers, 1994.

....approach. This approach, however, may cause some degradation in recognition performance. The use of incremental training (i.e. using only subsets of the training data at each iteration) is common in gradient based learning methods (e.g. back propagation training of connectionist systems [3, 4]) Recently, Neal and Hinton have discussed a theoretical justification for implementing an incremental E step for ML estimation [5] Restating their rationale: if the statistics for the E step are incrementally collected and the parameters are frequently estimated, it should speed the ....

Herv'e A. Bourlard and Nelson Morgan. Connectionist Speech Recognition: A Hybrid Approach. Kluwer Academic Publishers, Boston, MA, 1994.

....performance. The HMM training method presented in this paper is an incremental variant of the EM algorithm. A standard ML criterion is used, thus a prior distribution is not required. The use of incremental training is common in gradientbased learning methods (e.g. connectionist systems [3, 4]) and has recently been applied to gradient based training of HMMs [5] Neal and Hinton have discussed a theoretical justification for the incremental variants of the EM algorithm [6] If the statistics for the expectation step (E step) are incrementally collected and the parameters frequently ....

Herv'e A. Bourlard and Nelson Morgan. Connectionist Speech Recognition: A Hybrid Approach. Kluwer Academic Publishers, Boston, MA, 1994.

....algorithm. Fast convergence of an incremental generalized EM algorithm was also noted by Jordan and Jacobs in their work on hierarchical mixtures of experts [3] The use of incremental training is common in gradient based learning methods (e.g. back propagation training of connectionist systems [4]) and has recently been applied to gradient based training of HMMs [5] It was hoped that speed improvements could be obtained by applying a similar technique to the training of CD HMMs. 1 Partially funded by NSF grant MIP 9120843. 2 Partially funded by Wernicke ESPRIT Project (BRA 6487) This ....

Herv'e A. Bourlard and Nelson Morgan. Connectionist Speech Recognition: A Hybrid Approach. Kluwer Academic Publishers, Boston, MA, 1994.

....structure of human communication via spoken language, however, the decision process simply maps the rich continuous information into one of the discrete categories used in our language. We note that for ASR also, there are advantages to using continuous as opposed to discrete information [5, 7]. Finally, research has demonstrated that humans naturally integrate multiple sources of information in speech perception. Within the framework of the Fuzzy Logical Model of Perception, perceptual events are processed via three operations: feature evaluation, feature integration, and decision. The ....

H.A. Bourland & N. Morgan. Connectionist Speech Recognition: A Hybrid Approach Boston: Kluwer Academic Publishers, 1994.

No context found.

Herve Bourlard and Nelson Morgan. Connectionist Speech Recognition - A Hybrid Approach. Kluwer Academic Publishers, 1994.

....The database is phonetically hand transcribed. For the purposes of this study, we used what is known as the core subset : approximately two hours of the database for training and cross validation, and forty minutes (with non overlapping speakers) as a test set. We used ICSI s HMM MLP based [3] system. For our multiband system, we divided the frequency range into four bands of [216 778 Hz] 707 1631 Hz] 1506 2709 Hz] and [21213769 Hz] 1 , which roughly correspond to the formant regions. From the sub bands, we derived [3rd, 3rd, 2nd, 2nd] order RASTA PLP [7] features, respectively, ....

Herv e Bourlard and Nelson Morgan. Connectionist Speech Recognition -- A Hybrid Approach. Kluwer Academic Press, 1994.

No context found.

H. Bourland and N. Morgan. Connectionist Speech Recognition-A Hybrid Approach. Kluwer Academic, 1994.

No context found.

Herv'e Bourland and Nelson Morgan. 1994. Connectionist Speech Recognition: A Hybrid Approach. Kluwer Academic Publishers, Norwell, Mass.

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