Auto-encoder bottleneck features using deep belief networks

By Tara N. Sainath, Brian Kingsbury, Bhuvana Ramabhadran. 2012

187 citations recorded by the source corpus.

Subfields: Computer Vision and Pattern Recognition

How researchers see this paper

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

In Computer Vision and Pattern Recognition

Citations in the corpus

Extracting Domain Invariant Features by Unsupervised Learning for Robust Automatic Speech Recognition (2018, background): “Alternatively, robustness can also be achieved by training using features that are domain invariant [10, 11, 12, 13, 14].”
Improvements to Deep Convolutional Neural Networks for LVCSR (2013, background): “However, after features are extracted from both systems, the gains diminish down to 1-2% relative [18].”
Compressing deep neural networks using a rank-constrained topology (2015, methodology): “, through the use of bottle-neck layers [12, 13], or through a low-rank matrix factorization of the weights in the final layer of the DNN [14].”
Sequence classification using the high-level features extracted from deep neural networks (2014, unknown): “in speech feature coding and recognition [12][27].”
Deep Neural Networks for Single-Channel Multi-Talker Speech Recognition (2015, methodology): “We .rst review DNN based approaches to robust speech recognition in Section II.”
Deep Neural Networks for Acoustic Modeling in Speech Recognition (2012, background): “Recently, [62] investigated a less direct way of producing feature vectors for the GMM.”
Learning hidden unit contributions for unsupervised speaker adaptation of neural network acoustic models (2014, methodology): “In cases where speaker or channel adaptation can lead to significant improvements in accuracy, tandem systems often provide similar or reduced WERs compared with hybrid systems, owing to the applicability of GMM-based adaptation techniques [1, 9, 10].”
Multi-level adaptive networks in tandem and hybrid ASR systems (2013, methodology): “There are fewer comparisons between tandem and hybrid DNN systems: Sainath et al [21] found a tandem system using deep autoencoder bottleneck features was more accurate than a comparable hybrid system.”
Deep learning for limit order books (2016, unknown): “Various frameworks combining Gaussian mixtures with neural networks have been proposed by Variani et al. (2015), Demuynck and Triefenbach (2013), Paulik (2013), van den Oord and Schrauwen (2014), Yu and Seltzer (2011), Sainath et al. (2012), Deng and Chen (2014), and others.”
Improving deep neural networks for LVCSR using rectified linear units and dropout (2013, methodology): “Extracting bottleneck features [16] using deep nets trained with dropout would obviate most of the need for sequence training of the nets, but would abandon the goal of end-to-end training using a sequence level criterion.”
Convolutional Neural Networks for Distant Speech Recognition (2014, background): “Since the channels contain similar information (acoustic features shifted in time) we conjecture that the filter weights may be shared across different channels.”
Deep convolutional neural networks for LVCSR (2013, methodology): “Given this analysis, we then explore using CNNs on a 50-hr English Broadcast News (BN) task, in both hybrid [1] [5] [11] and neural network feature-based [12] setups.”
Deep neural network features and semi-supervised training for low resource speech recognition (2013, methodology): “As with traditional neural networks, DNNs have been used both as acoustic models that directly model context-dependent states of HMMs [17] and also as data-driven feature extractors [18, 19].”
Speech feature denoising and dereverberation via deep autoencoders for noisy reverberant speech recognition (2014, background): “Index Terms— robust speech recognition, feature denoising, denoising autoencoder, deep neural network”
Audio-visual speech recognition using deep learning (2014, methodology): “utilized a deep autoencoder as a dimensionality compression mechanism for self-organizing higher-level features from raw sensory inputs and utilized the acquired higher-level features as inputs to a conventional GMM-HMM system [43].”