Improved Bottleneck Features Using Pretrained Deep Neural Networks

By Dong Yu, M. Seltzer. 2011

314 citations recorded by the source corpus.

Subfields: Artificial Intelligence, Representation & Self-Supervised Learning, 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

1. Dong Yu
2. Li Deng
3. Jinyu Li
4. James R. Glass
5. Yoshua Bengio
6. Tara N. Sainath
7. Yu Zhang
8. Shinji Watanabe

In Representation & Self-Supervised Learning

1. James R. Glass
2. Hung-yi Lee
3. Jinyu Li
4. Thomas Hain
5. Wei-Ning Hsu
6. Dong Yu
7. Yu-An Chung
8. Yu Zhang

In Signal Processing

1. Dong Yu
2. Jinyu Li
3. Li Deng
4. Tara N. Sainath
5. Yifan Gong
6. James R. Glass
7. Florian Metze
8. Shinji Watanabe

Citations in the corpus

• Unsupervised Speech Representation Learning Using WaveNet Autoencoders (2019, background): “Similarly, the speech community has used bottleneck features extracted from networks trained on phoneme prediction tasks [30], [31] as feature representations for speech recognition systems.”
• Deep maxout networks for low-resource speech recognition (2013, methodology): “In tandem systems, we use DNNs to generate phone posteriors or bottleneck features (BNF), and build normal GMMHMM models with the discriminative front-end [3, 4, 5, 6].”
• Restructuring of deep neural network acoustic models with singular value decomposition (2013, methodology): “The work in [10] shows that DNN trained bottleneck feature reduces word error rate by 16% relatively on a large vocabulary business search task”
• An empirical evaluation of zero resource acoustic unit discovery (2017, unknown): “To date language-independent bottleneck (BN) features have been demonstrated as effective speech representations in improving ASR accuracies [6, 7].”
• Boosting attribute and phone estimation accuracies with deep neural networks for detection-based speech recognition (2012, methodology): “We extend the conventional shallow MLPs used in [12] to deep neural networks (DNNs) [13], which has been shown to have very good theoretical properties [14] and demonstrated superior performances for both phone [16][17] and word recognition [15][13][18][19].”
• Deep neural network features and semi-supervised training for low resource speech recognition (2013, methodology): “In separate experiments, placing the bottleneck layer in the center as used in [18] did not provide any gains.”
• Deep Neural Networks for Single-Channel Multi-Talker Speech Recognition (2015, background): “layer, or bottleneck feature [26], [27] if it is extracted from a”
• Hierarchical Bayesian combination of plug-in maximum a posteriori decoders in deep neural networks-based speech recognition and speaker adaptation (2017, methodology): “The SMAP (Shinoda and Lee, 2001) technique is used to perform speaker adaptation within BN-CD-GMM-HMMs.”
• Expressive TTS Training With Frame and Style Reconstruction Loss (2020, background): “other layers, it creates a constriction in the network that forces the information pertinent to emotion classification into a low dimensional prosody representation [91].”
• An introduction to computational networks and the computational network toolkit (invited talk) (2014, background): “There is a common property in key machine learning models, such as deep neural networks (DNNs) [2, 3, 4, 5, 6, 7], convolutional neural networks (CNNs) [8, 9, 10, 11, 12, 13, 14, 15, 16, 17], and recurrent neural networks (RNNs) [18, 19, 20, 21, 22].”
• Deep neural networks employing Multi-Task Learning and stacked bottleneck features for speech synthesis (2015, unknown): “Bottleneck features have been extensively used in speech recognition [18, 19, 20].”
• Using neural network front-ends on far field multiple microphones based speech recognition (2014, background): “It is possibly due to the limited number of units which creates a constriction in the network and further forces the information pertinent to classification into a low dimensional representation [12, 13, 14].”
• A scalable approach to using DNN-derived features in GMM-HMM based acoustic modeling for LVCSR (2013, methodology): “More recently, encouraged by the success of DNN-HMM approach, several extensions to the conventional tandem and bottleneck-feature approaches are proposed, which include using a bottleneck in a pre-trained DNN in [9], and using auto-encoder bottleneckfeatures on top of a DNN in [10].”
• Investigation on dimensionality reduction of concatenated features with deep neural network for LVCSR systems (2012, background): “So, PCA fails to show us any satisfying results in our experiments because of the limitation of its linearity.”
• Deep learning for limit order books (2016, methodology): “Various frameworks combining Gaussian mixtures with neural networks have been proposed by Variani et al. (2015), Demuynck and Triefenbach (2013), Paulik (2013), van den Oord and Schrauwen (2014), Yu and Seltzer (2011), Sainath et al. (2012), Deng and Chen (2014), and others.”
• Integrating Deep Neural Networks into Structural Classification Approach based on Weighted Finite-State Transducers (2012, methodology): “Since it has been reported that the pretraining techniques developed for DNNs are also valid for extracting efficient bottleneck features [11], we can realize an effective combination of these two methods by employing these techniques.”
• Extracting deep bottleneck features using stacked auto-encoders (2013, background): “for extracting bottleneck features, with the bottleneck being a small RBM placed in the middle of the network [12].”
• Exploiting deep neural networks for detection-based speech recognition (2013, background): “In detection-based ASR paradigms (e.g., [11]), a set of neural networks learns the mapping from a spectral-based feature space to a phonetic feature space.”