Learning hidden unit contributions for unsupervised speaker adaptation of neural network acoustic models

By Pawel Swietojanski, S. Renals. 2014

246 citations recorded by the source corpus.

Subfields: Artificial Intelligence, 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

In Signal Processing

Citations in the corpus

Personalization of End-to-End Speech Recognition on Mobile Devices for Named Entities (2019, background): “There are existing speaker adaptation techniques for neural networks that offer a more compact model representation by having a large number of global parameters shared across different speakers and a smaller number of speaker-specific parameters [13, 14, 15, 16, 17, 18, 19, 20].”
Continual Learning with Adaptive Weights (CLAW) (2019, background): “The variable aT is an unconstrained adaptation value, similar to that in (Swietojanski & Renals, 2014).”
An analysis of environment, microphone and data simulation mismatches in robust speech recognition (2017, unknown): “This is widely recognized as good practice and many solutions are available to improve robustness to this mismatch (Gales, 1998; Shinoda, 2011; Karaﬁ´at et al., 2011; Swietojanski and Renals, 2014).”
Adapting and controlling DNN-based speech synthesis using input codes (2017, unknown): “Wu et al. [13] proposed speaker adaptation using i-vectors as input, or by adapting hidden unit contributions (LHUC [14]), or by applying output transforms deﬁned by GMMs, or combinations of these.”
Improving the interpretability of deep neural networks with stimulated learning (2015, background): “One example of this is in Learning Hidden Unit Contribution (LHUC) [14] where the number of parameters to be adapted is equal to the number of neurons in the hidden layer.”
Spoken Content and Voice Factorization for Few-Shot Speaker Adaptation (2020, background): “Many researchers have tried to make the acoustic model more interpretable so as to adapt new speaker with small hypothesis space [5–7].”
Machine Speech Chain with One-shot Speaker Adaptation (2018, methodology): “Then, they performed model adaptation using the learning hidden unit contributions at the middle level based on the speaker dependent parameters [23].”
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”
Embedding-Based Speaker Adaptive Training of Deep Neural Networks (2017, methodology): “In another effort under the name of learning hidden unit contributions (LHUC), only the amplitudes of the outputs of hidden units are re-scaled with the adaptation data [10].”
Multilingual Sequence-to-Sequence Speech Recognition: Architecture, Transfer Learning, and Language Modeling (2018, background): “The ‡ All three authors share equal contribution authors also use linear hidden unit contribution (LHUC) [13] technique to rescale the hidden unit outputs for each language as a way to adapt to a particular language.”
On training the recurrent neural network encoder-decoder for large vocabulary end-to-end speech recognition (2016, background): “The past few years have seen signiﬁcant advancements in speech recognition based on this hybrid architecture including using different neural network architectures [4, 5, 6], sequence training [7, 8, 9] and speaker adaptation [10, 11, 12].”
Low-rank plus diagonal adaptation for deep neural networks (2016, methodology): “A DNN has been adapted by adjusting amplitude, bias and slopes in the node activation functions, such as learning hidden unit contribution (LHUC) [13] and sigmoid adaptation [14].”
Recent progresses in deep learning based acoustic models (2017, background): “For example, when the adaptation data is very limited, the kXk matrix can be reduced to a diagonal matrix, such as learning hidden unit contribution (LHUC) [103], [104] and sigmoid adaptation [105].”
Linear Networks Based Speaker Adaptation for Speech Synthesis (2018, methodology): “I-vectors [13] are augmented with linguistic features to represent speaker identity information at input level, and the Learning Hidden Unit Contributions (LHUC) [14] is used to conduct model adaptation at the middle level.”
Factorized Hidden Layer Adaptation for Deep Neural Network Based Acoustic Modeling (2016, methodology): “Another effective model adaptation technique is known as LHUC [42], [43], which learns speaker dependent hidden unit contributions during adaptation.”
Maximum a posteriori adaptation of network parameters in deep models (2015, methodology): “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”
Speaker adaptive training in deep neural networks using speaker dependent bottleneck features (2016, methodology): “In learning hidden unit contributions (LHUC), an SD vector is attached to every hidden layer learnt from the test speaker, which are applied to DNN hidden units with element wise multiplication [12].”
Dynamic Layer Normalization for Adaptive Neural Acoustic Modeling in Speech Recognition (2017, methodology): “In [10, 11], a speaker independent model was adapted to a specific speaker with speaker dependent parameters at each hidden layer.”