Tandem connectionist feature extraction for conventional HMM systems

By Hynek Hermansky, D. Ellis, Sangita Sharma. 2000

832 citations recorded by the source corpus.

Subfields: Artificial Intelligence, Signal Processing

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Signal Processing

Citations in the corpus

Deep Learning: Methods and Applications (2014, methodology): “[154], has been adopted, where the output of the neural networks in the form of posterior probabilities for phone classes, are used, often in conjunction with the acoustic features to form new augmented input features, in a separate GMM–HMM system.”
Understanding how Deep Belief Networks perform acoustic modelling (2012, background): “Some ASR systems use feedforward neural networks [1, 2].”
Speech recognition and keyword spotting for low-resource languages: Babel project research at CUED (2014, methodology): “The DNN can be used as feature extractor for a standard GMM-based HMM system [5, 6], this approach is referred to as Tandem.”
Context-dependent deep neural networks for commercial Mandarin speech recognition applications (2013, methodology): “As a result, the main practical contribution of ANNs was to provide useful features, namely tandem or bottleneck features [7], in which the posterior probability of each phone was estimated using ANN.”
Hierarchical bottle neck features for LVCSR (2010, background): “Nevertheless, short-term features like PLPs have been used as input features as well [7].”
RASR/NN: The RWTH neural network toolkit for speech recognition (2014, background): “The available training criteria are cross entropy, squared error, and binary divergence [12].”
Factorized Hidden Layer Adaptation for Deep Neural Network Based Acoustic Modeling (2016, background): “There have been a number of attempts to combine GMM and DNN to leverage on the reliable adaptation performance of the former and high quality classification performance of the latter.”
A scalable approach to using DNN-derived features in GMM-HMM based acoustic modeling for LVCSR (2013, methodology): “For example, a so-called tandem approach [7] uses log-posterior probabilities generated by the softmax output layer of a multi-layer perceptron (MLP), or the weighted sums of the outputs of the”
Investigation of unsupervised adaptation of DNN acoustic models with filter bank input (2014, background): “The second model configuration, which is called a tandem, exploits the DNNs to perform nonlinear discriminative feature transformation [6, 7].”
The language-independent bottleneck features (2012, methodology): “Our work is situated in the Tandem LVCSR framework [5], where a traditional Hidden Markov Model (HMM) based LVCSR system processes features generated by ANNs.”
Sequence classification using the high-level features extracted from deep neural networks (2014, methodology): “The “tandem” approach makes use of the log-posterior probabilities generated by the softmax output layer of a shallow neural network [18].”
Lattice-based optimization of sequence classification criteria for neural-network acoustic modeling (2009, background): “Directions for future work include evaluation of other sequence classification criteria [11] for neural network training, comparison of frame-based training and sequence-based training in tandem systems [17], and development of methods for neural network training that scale to larger networks and da”
End-to-End Speech Recognition: A Survey (2023, background): “…1 The term “classical” here refers to the former, long-term, state-of-the-art ASRarchitecturebasedonthedecompositionintoacousticandlanguagemodel,andwithacousticmodelingbasedonhiddenMarkovmodels. the acoustic feature set (e.g., non-linear discriminant/tandem approach [6], [7]).”
Enhanced Phone Posteriors for Improving Speech Recognition Systems (2010, methodology): “…a particular state of the set of possible HMM states, it is possible to train the MLP in a discriminative way, to generate posterior probabilities of the output classes conditioned on the input, i.e., , where is a spectral feature frame at time , and represents the event of having phone at time .”
Audio-visual speech recognition using deep learning (2014, methodology): “As an experimental study for an AVSR task, we adopted a rather simple tandem approach, a connectionist-HMM [16].”