Joint training of convolutional and non-convolutional neural networks

By Hagen Soltau, George Saon, Tara N. Sainath. 2014

83 citations recorded by the source corpus.

Subfields: Artificial Intelligence, Model Architectures, Optimization & Training, 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 Model Architectures

In Optimization & Training

In Signal Processing

Citations in the corpus

Deep Speech 2 : End-to-End Speech Recognition in English and Mandarin (2015, background): “er to reduce the number of time-steps [26]. Convolutions in frequency and time domains, when applied to the spectral input features prior to any other processing, can slightly improve ASR performance [1, 53, 60]. Convolution in frequency attempts to model spectral variance due to speaker variability”
Wav2Letter: an End-to-End ConvNet-based Speech Recognition System (2016, background): “The improvements brought by deep neural networks (DNNs) [14, 10] and convolutional neural networks (CNNs) [24, 25] for acoustic modeling only extend this training pipeline.”
Deep Speech: Scaling up end-to-end speech recognition (2014, background): “N-HMM FSH) [28] 16.0 23.7 19.9 Seide et al. (CD-DNN) [39] 16.1 n/a n/a Kingsbury et al. (DNN-HMM sMBR HF) [22] 13.3 n/a n/a Sainath et al. (CNN-HMM) [36] 11.5 n/a n/a Soltau et al. (MLP/CNN+I-Vector) [40] 10.4 n/a n/a Deep Speech SWB 20.0 31.8 25.9 Deep Speech SWB + FSH 12.6 19.3 16.0 Table 3: Publi”
Letter-Based Speech Recognition with Gated ConvNets (2017, background): “The performance improvements brought by deep neural networks (DNNs) (Mohamed et al., 2012; Hinton et al., 2012) and convolutional neural networks (CNNs) (Sercu et al., 2016; Soltau et al., 2014) for acoustic modeling only extend this training pipeline.”
English Conversational Telephone Speech Recognition by Humans and Machines (2017, background): “In [7], we have shown that convolutional and nonconvolutional AMs have comparable performance and good complementarity.”
The IBM 2015 English conversational telephone speech recognition system (2015, methodology): “We have presented a set of improvements to our English Switchboard system that lowered the error rate substantially compared to our previous best result [13].”
On speaker adaptation of long short-term memory recurrent neural networks (2015, background): “Although making significant advances, DNNs, as well as the more advanced convolutional neural networks (CNNs) [4, 5, 6, 7], can model only limited temporal dependency that is provided by the fixed-size window of acoustic frames.”
Efficient Knowledge Distillation from an Ensemble of Teachers (2017, background): “These strategies include combining various acoustic feature sets together at the input of acoustic models, joint training of complex acoustic models after fusing various neural network architectures [1, 2] and combination of acoustic scores predicted at the output of various acoustic models.”
Data Augmentation for Deep Neural Network Acoustic Modeling (2015, methodology): “2 shows a mean and speaker normalized LogMel feature extraction pipeline used in IBM CNN systems [31][32].”
Dense Prediction on Sequences with Time-Dilated Convolutions for Speech Recognition (2016, methodology): “SWB CH XE ST XE ST Classic 512 CNN [18] 12.”
Wise teachers train better DNN acoustic models (2016, methodology): “DNN and CNN were jointly trained in [50] for a LVCSR task.”
Very deep convolutional neural networks for robust speech recognition (2016, background): “Considering that topographical features, such as FBANK, are more appropriate than the non-topographical feature, such as fMLLR and i-vector, for the CNN usage [29], an auxiliary feature joint training framework is proposed within the very deep CNN architecture, shown as the left part of Figure 2.”
Advances in Very Deep Convolutional Neural Networks for LVCSR (2016, methodology): “We performed stochastic sequence training as opposed to Hessian-Free sequence training [29] which was used in our baselines [6, 7].”
Speaker Adaptive Training of Deep Neural Network Acoustic Models Using I-Vectors (2015, methodology): “Also, in addition to DNNs, our framework can be applied readily to other types of acoustic models such as convolutional neural networks (CNNs) [22], [23], [24], [25] and recurrent neural networks (RNNs) [26], [27], [28], [29], [30].”
Distilling Knowledge from Ensembles of Neural Networks for Speech Recognition (2016, methodology): “Examples include “stacking” posterior predictions from a deep neural network (DNN), convolutional neural network (CNN), recurrent neural network (RNN) [3] and hybrid architectures that fuse CNN and DNN structures earlier in the model [4, 5].”
Deep Scattering Spectrum with deep neural networks (2014, methodology): “Hence a CNN/DNN combination network was used to process these features, which has been explored before for combining multiple feature streams [21, 22].”
A time delay neural network architecture for efficient modeling of long temporal contexts (2015, background): “3 CNN/DNN joint training + fMLLR features + iVectors [25] - 12.”