Size matters: an empirical study of neural network training for large vocabulary continuous speech recognition

By D. Ellis, Nelson Morgan. 1999

Subfields: Optimization & Training, Speech & Audio

How researchers see this paper

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

In Optimization & Training

1. Bryan Catanzaro
2. Andrew L. Maas
3. Andrew Y. Ng
4. Adam Coates
5. Hynek Hermansky
6. Awni Y. Hannun
7. Dan Jurafsky
8. Ziang Xie

In Speech & Audio

1. Awni Y. Hannun
2. Bryan Catanzaro
3. Gabriel Synnaeve
4. Frantisek Grézl
5. Jinyu Li
6. Ryan J. Prenger
7. Hervé Bourlard
8. Eric Battenberg

Citations in the corpus

• Deep Speech 2 : End-to-End Speech Recognition in English and Mandarin (2015, methodology): “sers in Section 7. 2 Related Work This work is inspired by previous work in both deep learning and speech recognition. Feed-forward neural network acoustic models were explored more than 20 years ago [7, 50, 19]. Recurrent neural networks and networks with convolution were also used in speech recogn”
• Deep Speech: Scaling up end-to-end speech recognition (2014, background): “Neural network acoustic models and other connectionist approaches were first introduced to speech pipelines in the early 1990s [1, 34, 11].”
• Building DNN acoustic models for large vocabulary speech recognition (2014, background): “Indeed, training time of neural networks has been a persistent issue throughout history, researchers often utilized whatever specialized computing hardware was available at the time [54], [55].”
• Programmable Neurocomputing (background): “For example, computations may require greater than 10 15 arithmetic operations and operate on data sets containing several gigabytes of data [5].”
• Analysis of MLP-Based Hierarchical Phoneme Posterior Probability Estimator (2011, methodology): “Consequently, the performance of AS R systems using MLP acoustic models can be improved using the following three broad strategies: (a) using richer acou stic features (b) increasing the capacity of the MLP (but this approach is often limited by the amount of training data [16] ) and (c) using finer”
• Lattice-based optimization of sequence classification criteria for neural-network acoustic modeling (2009, background): “(3)”
• Probabilistic and Bottle-Neck Features for LVCSR of Meetings (2007, background): “The input vector is concatenation of all band-conditioned outputs and sufficient size of the hidden layer is also needed as it affects the final classification accuracy [10].”