Adaptation of context-dependent deep neural networks for automatic speech recognition

By Kaisheng Yao, Dong Yu, Frank Torsten Bernd Seide, Hang Su, Li Deng, Yifan Gong. 2012

215 citations recorded by the source corpus.

Subfields: Artificial Intelligence, Speech & Audio, 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. Jinyu Li
3. Li Deng
4. Yoshua Bengio
5. Shinji Watanabe
6. James R. Glass
7. Jianfeng Gao
8. Yifan Gong

In Speech & Audio

1. Dong Yu
2. Jinyu Li
3. Li Deng
4. Shinji Watanabe
5. Daniel Povey
6. Florian Metze
7. Sanjeev Khudanpur
8. Yoshua Bengio

In Signal Processing

1. Dong Yu
2. Jinyu Li
3. Li Deng
4. Yifan Gong
5. Shinji Watanabe
6. Khe Chai Sim
7. S. Renals
8. Florian Metze

Citations in the corpus

• Unsupervised Pre-Training of Bidirectional Speech Encoders via Masked Reconstruction (2020, methodology): “For domain adaptation we use linear input networks (LIN, [29, 30]), which inserts an additional linear layer (initialized as the identity mapping) between the input and the”
• Recent advances in convolutional neural networks (2015, background): “s) in large vocabulary continuous speech recognition (LVCSR) and obtained encouraging results [300{304], however, their networks are susceptible to performance degradations under mismatched condition [305], such as dierent recording conditions etc . CNNs have shown better performance over GMM-HMMs a”
• Deep Learning: Methods and Applications (2014, background): “A number of exploratory studies on DNN adaptation made use of the same or related linear transformation methods [223, 401, 402].”
• Speaker adaptive training in deep neural networks using speaker dependent bottleneck features (2016, background): “These layers can be positioned either towards the front [8, 9], middle [9] or towards the output layer [10].”
• KL-divergence regularized deep neural network adaptation for improved large vocabulary speech recognition (2013, methodology): “can compare the KLD-Reg approach with fDLR [25], which performs slightly better than oDLR [20] on the same dataset.”
• Learning hidden unit contributions for unsupervised speaker adaptation of neural network acoustic models (2014, methodology): “Smaller subsets of the DNN weights may be modified, including adaptation of output layer biases [20], and adaptation of the bias and slope of hidden units (a contrast experiment in [29]).”
• Cluster adaptive training for deep neural network (2015, background): “Linear transforms are also widely used to adapt feature or hidden-layer output [9, 10, 11, 12].”
• On Multi-Domain Training and Adaptation of End-to-End RNN Acoustic Models for Distant Speech Recognition (2017, background): “data, keeping the rest of the network unchanged [22, 23].”
• Low-rank plus diagonal adaptation for deep neural networks (2016, methodology): “A simple and popular approach to combat the mismatch problem is to adapt the DNN model by applying a linear transformation on either input features, hidden layers, or the input to the softmax layer [5, 6, 7, 8, 9].”
• Speaker adaptation of context dependent deep neural networks based on MAP-adaptation and GMM-derived feature processing (2014, methodology): “…system: to the input features, as in Linear Input Network transformation (LIN) [7]-[10] or Feature-space Discriminative Linear Regression (fDLR) [5][11]; to the activations of hidden layers, as in Linear Hidden Network transformation (LHN) [7][8]; or to the softmax layer, as in LON [9] or in…”
• Fast Adaptation of Deep Neural Network Based on Discriminant Codes for Speech Recognition (2014, background): “On the other hand, maximum likelihood linear regression (MLLR) [3] and con­strained MLLR (CMLLR) [4] have been reported to work much better when only a small amount of adaptation data is used.”
• Rapid adaptation for deep neural networks through multi-task learning (2015, background): “In recent years, several adaptation approaches have been proposed to address this mismatch issue, such as regularization based [3, 4], subspace based [5, 6], transformation based [7, 8, 9, 10], i-Vector based [11], native neural network adaptation [12, 13, 14], factorized adaptation [15] and speaker”
• Speaker adaptation of neural network acoustic models using i-vectors (2013, background): “Alternatively, [2] have looked at adapting only the biases.”
• Recent advances in deep learning for speech research at Microsoft (2013, background): “We have recently investigated the affine transformation and several of its variants for adaptation of the top hidden layer [52].”
• Maximum a posteriori adaptation of network parameters in deep models (2015, background): “Recent approaches, such as regularization based [7, 8], subspace based [9, 10], transformation based [6, 11, 12, 13], i-Vector based [14], native neural network based [15, 16, 17], factorization based [18] and fast adaptation schemes based on discriminant speaker codes [19, 16, 20, 21], have been pr”
• Embedding-Based Speaker Adaptive Training of Deep Neural Networks (2017, methodology): “Or, alternatively, the input or output layer is adapted by an affine transformation estimated using the adaptation data [6][7][8].”
• Factorized Hidden Layer Adaptation for Deep Neural Network Based Acoustic Modeling (2016, methodology): “Moreover, some approaches perform the adaptation on a subset of parameters, including the last hidden layer [45], [46], input layer [46], biases [47], output layer biases [48], or more active hidden units of the network [45].”