Phoneme Recognition Using Time-Delay Neural Networks

Waibel, Alexander; Hanazawa, Toshiyuki; Hinton, Geoffrey E.; Shikano, Kiyohiro; Lang, Kévin

doi:10.1016/b978-0-08-051584-7.50037-1

Cited by 644 publications

(733 citation statements)

References 13 publications

Supporting

Mentioning

712

Contrasting

Unclassified

Order By: Relevance

“…In the domain of speech recognition Time Delay Neural Networks (TDNNs) have been proven to be powerful tools for various segmentation and classification problems (Waibel et al, 1989), overcoming also the problem of time variations in the input signal. For this reason the application of TDNNs was considered for the solution of the analysis problem in the RPD system.…”

Section: Discussionmentioning

confidence: 99%

Untitled

Friedrich¹,

Münch²,

Dillmann³

et al. 1996

Machine Learning

View full text Add to dashboard Cite

Abstract. Programming by Demonstration (PbD) is a programming method that allows software developers to add new functionalities to a system by simply showing them in the form of few examples. In the robotics domain it has the potential to reduce the amount of time required for programming and also to make programming more "natural". Just imagine the task of assembling a torch by a manipulator. Wouldn't it be nice to just assemble the torch with one's own hands, watched by video and laser cameras and maybe wearing data gloves, i. e. sensors that provide the data to automatically generate the necessary robot program for the assembly task? And wouldn't it be even nicer to demonstrate the task with few different torches, but achieving an assembly function for all possible variants of them? In order to realize such a PbD environment, at least two major problems have to be solved. First, the sensor data have to be transformed to high-level situation-action descriptors, a task that is not yet solved in general. Second, if a generalization is required, induction algorithms must be applied to the recorded and transformed traces, aiming to find the most general user-intended function from only few examples. In this article we will concentrate only on the second problem. The described experimental environment consists of an industrial robot (PUMA 260b), a 6D teach mouse input device, and some sensors. Various data can be recorded during a demonstration for further processing in the PbD system running on a workstation. The objective is to explore the possibilities of integrating learning and clustering algorithms for automated robot programming. In particular it is investigated how human interaction within system-as well as user-initiated dialogs can support the induction component.

show abstract

Section: Discussionmentioning

confidence: 99%

Untitled

Friedrich¹,

Münch²,

Dillmann³

et al. 1996

Machine Learning

View full text Add to dashboard Cite

show abstract

“…with the t index in (9). Therefore, for MLP with IIR synapses, expression (8) means that each back propagating error at layer l is a summation of all the delta's at the following layer filtered by the non-causal version of the respective IIR filter. The non-causal version of each filter can be obtained practically, convoluting with time reverted impulse response or time reverting the output of the filter giving the input in a reversed time scale.…”

Section: A Unifying Formulation For Learning Methods In Locally Recurmentioning

confidence: 99%

“…However it can only be applied to the non-recurrent MLP with Finite Impulse Response (FIR) filter synapses [3,7,21,9], i.e. FIR-MLP often called Time Delay Neural Network (TDNN) [8].…”

Section: Introductionmentioning

confidence: 99%

A Unifying View of Gradient Calculations and Learning for Locally Recurrent Neural Networks

Campolucci

Uncini

Piazza

1998

Neural Nets WIRN VIETRI-97

View full text Add to dashboard Cite

In this paper a critical review of gradient-based training methods for recurrent neural networks is presented including Back Propagation Through Time (BPTT), Real Time Recurrent Learning (RTRL) and several specific learning algorithms for different locally recurrent architectures. From this survey it comes out the need for a unifying view of all the specific procedures proposed for networks with local feedbacks, that keeps into account the general framework of recurrent networks learning: BPTT and RTRL. Therefore a learning method for local feedback network is proposed which combines together the best feature of BPTT, i.e. the lowest complexity, and of RTRL, i.e. the on-line operation, and includes as special case several specific algorithms already proposed, such as Temporal Back Propagation, Back Propagation for Sequences, Back-Tsoi algorithm and some others. In the general version, this new training method allows on-line efficient and accurate gradient calculation. It compares favourably with the previous algorithms in stability, speed/complexity trade off, accuracy.

show abstract

“…The feature vectors computed by the feature extractor must be classified as (the produced) visemes. In order to incorporate temporal aspects in the classification process a Time-delay Neural Network (TDNN) [3,5], was used to classify the feature vectors.…”

Section: A Prototype Of the Lip Reading Systemmentioning

confidence: 99%

Classifying Visemes for Automatic Lipreading

Visser

Poel

Nijholt

1999

Lecture Notes in Computer Science

View full text Add to dashboard Cite

Abstract. Automatic lipreading is automatic speech recognition that uses only visual information. The relevant data in a video signal is isolated and features are extracted from it. From a sequence of feature vectors, where every vector represents one video image, a sequence of higher level semantic elements is formed. These semantic elements are "visemes" the visual equivalent of "phonemes" The developed prototype uses a Time Delayed Neural Network to classify the visemes.

show abstract

Phoneme Recognition Using Time-Delay Neural Networks

Cited by 644 publications

References 13 publications

Untitled

Untitled

A Unifying View of Gradient Calculations and Learning for Locally Recurrent Neural Networks

Classifying Visemes for Automatic Lipreading

Contact Info

Product

Resources

About