October 19, 2012IEEE Signal Processing Magazine
Deep Neural Networks for Acoustic Modeling in Speech Recognition: The Shared Views of Four Research Groups
Authors
Discussion
Loading...
Member takes
Implication
Collaborative review reveals deep neural networks outperform Gaussian mixture models in speech recognition, highlighting a major advance in acoustic modeling.
Key Points
- To present a unified review from four research groups assessing the efficacy of deep neural networks as replacements for Gaussian mixture models in speech acoustic modeling.
- Synthesized empirical results across four research groups evaluating deep feed-forward neural networks containing multiple hidden layers for acoustic modeling.
- Evaluated architectures taking multiple frames of speech coefficients as input to generate posterior probabilities over hidden Markov model states, replacing traditional Gaussian mixture models.
- Deep neural networks consistently outperformed standard Gaussian mixture models across multiple benchmark speech recognition tasks, frequently by large margins.
- Novel training methods for multi-layered feed-forward networks enabled substantially improved modeling of acoustic input when integrated with temporal hidden Markov models.
Cite This Study
Hinton et al. (2012) studied this question.