Constructing long short-term memory based deep recurrent neural networks for large vocabulary speech recognition

By Xiangang Li, Xihong Wu. IEEE International Conference on Acoustics, Speech, and Signal Processing, 2014

324 citations recorded by the source corpus.

Long short-term memory (LSTM) based acoustic modeling methods have recently been shown to give state-of-the-art performance on some speech recognition tasks. To achieve a further performance improvement, in this research, deep extensions on LSTM are investigated considering that deep hierarchical model has turned out to be more efficient than a shallow one. Motivated by previous research on constructing deep recurrent neural networks (RNNs), alternative deep LSTM architectures are proposed and empirically evaluated on a large vocabulary conversational telephone speech recognition task. Meanwhile, regarding to multi-GPU devices, the training process for LSTM networks is introduced and discussed. Experimental results demonstrate that the deep LSTM networks benefit from the depth and yield the state-of-the-art performance on this task.

Subfields: Artificial Intelligence, Signal Processing

View Paper: https://www.semanticscholar.org/paper/1250ef2f19d0ac751ec7d0e2a22e741ecb40ea92

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

Building DNN acoustic models for large vocabulary speech recognition (2014, background): “A deep understanding of these feedforward DNNs can help guide the emerging work utilizing a wide variety of new architectures based on recurrent neural networks (Graves et al., 2013; Li and Wu, 2015; Sak et al., 2014; Vinyals et al., 2012; Weng et al., 2014), sequence-discriminative training (Kingsb”
DeepBach: a Steerable Model for Bach Chorales Generation (2016, background): “g. 4a a neural network with one hidden layer of size 200 and ReLU [25] nonlinearity and as the \Stacked LSTMs brick" two LSTMs on top of each other, each one being of size 200 (see Fig. 2 (f) in [21]). We experimentally found that adding dropout or sharing weights between the embedding layers i”
A Survey of Machine Learning Techniques Applied to Software Defined Networking (SDN): Research Issues and Challenges (2019, background): “Long Short-Term Memory (LSTM) [101], [102] is the most commonly-used type of recurrent NNs, which has a good ability to capture long-term dependencies.”
Point2Sequence: Learning the Shape Representation of 3D Point Clouds with an Attention-based Sequence to Sequence Network (2018, unknown): “To aggregate sequential feature vectors, recurrent neural networks (Elman 1990) have shown preeminent performance in many popular tasks such as speech recognition (Li and Wu 2015) or handwriting recognition (Bertolami et al. 2009).”
On speaker adaptation of long short-term memory recurrent neural networks (2015, methodology): “Past work [12, 13, 14, 15] has applied LSTM-RNNs to acoustic modeling and shown state-of-the-art performance.”
A comprehensive study of deep bidirectional LSTM RNNS for acoustic modeling in speech recognition (2016, background): “[3, 4, 10, 11, 12, 13, 14, 15] investigate various bidirectional and unidirectional LSTM topologies with optional projection in some cases combined with convolutional or feed-forward layers for acoustic modeling in ASR.”
Modeling speaker variability using long short-term memory networks for speech recognition (2015, methodology): “Thus, in this paper, we have 5529 senones against 3302 senones in [36], leading to slightly better experimental results than that in [36].”
Anomaly Detection in Cyber Physical Systems Using Recurrent Neural Networks (2017, background): “In addition, stacking recurrent hidden layers in a neural networks has shown to capture the structure of the time series [10].”
A Survey on Deep Learning (2018, background): “Therefore, Long Short-Term Memory (LSTM) [93] is utilized to handle this issue by providing memory blocks in its recurrent connections.”
Dataset Augmentation in Feature Space (2017, methodology): “Dai and Le follow the original seq2seq approach of Sutskever et al. (2014) and use the context vector as input to the decoder only on the ﬁrst time step, then use the output of the previous times step as inputs for all subsequent time steps as follows: where f is the LSTM function, s is the state of”