Optimizing bottle-neck features for lvcsr

By Frantisek Grézl, Petr Fousek. 2008

146 citations recorded by the source corpus.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, 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 Computer Vision and Pattern Recognition

In Signal Processing

Citations in the corpus

Deep Learning: Methods and Applications (2014, neutral): “Several variants of the DNN-based bottleneck-feature approach have been explored; see details in [16, 137, 201, 285, 308, 368].”
TED-LIUM: an Automatic Speech Recognition dedicated corpus (2012, methodology): “The acoustic models were trained in the same manner, to the exception that we added a multi-layer perceptron (MLP) using the Bottle-Neck feature extraction as described in (Grézl and Fousek, 2008).”
On the compression of recurrent neural networks with an application to LVCSR acoustic modeling for embedded speech recognition (2016, methodology): “, by introducing bottleneck layers [18] or through a low-rank matrix factorization layer [19].”
Auto-encoder bottleneck features using deep belief networks (2012, background): “1The authors are aware that [9] proposed a BN architecture which provided gains over regular speech features when not used in tandem.”
Extracting deep bottleneck features using stacked auto-encoders (2013, methodology): “However, the difference between DBNFs from pre-trained and purely supervised trained networks was smaller in terms of CER.”
Using neural network front-ends on far field multiple microphones based speech recognition (2014, background): “Index Terms— speech recognition, multiple microphone, beamforming, deep neural networks”
Transfer Learning for Speech Recognition on a Budget (2017, methodology): “The method presented in this paper differs from Vu and Schultz (2013) in that it does not force the representation to be compressed into bottleneck features (Grezl and Fousek, 2008) and use the result as the output of the pre-trained network.”
Backpropagation training for multilayer conditional random field based phone recognition (2010, methodology): “For example, Bottleneck Features [13] utilize a narrow hid- den layer in a 5-layer MLP to compress phonetic information into a set of features usable in the Tandem MLP-HMM setting.”
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].”
Making Deep Belief Networks effective for large vocabulary continuous speech recognition (2011, methodology): “Alternatively, approaches such as TANDEM [3] and bottleneck features [4] derive a set of features from the MLP, which are then used as input features into a traditional GMM/HMM system.”
Improved Bottleneck Features Using Pretrained Deep Neural Networks (2011, methodology): “When generating neural-network-based features, the input is typically generated using variants of TRAPS features, in which long temporal windows of critical band energies are processed [9][10].”
Hierarchical bottle neck features for LVCSR (2010, methodology): “Generally, a 3-layer NN structure is used but other topologies have been investigated as well [8, 10].”
Context-Dependent Pre-Trained Deep Neural Networks for Large-Vocabulary Speech Recognition (2012, background): “Although a neural network nominally containing three hidden layers (the largest number of layers investigated in [55]) might be used to create bottle-neck features, if the feature layer is the middle hidden layer then the resulting features are only produced by an encoder with a single hidden layer.”