Unsupervised cross-lingual knowledge transfer in DNN-based LVCSR

By Pawel Swietojanski, Arnab Ghoshal, S. Renals. 2012

148 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

1. Dong Yu
2. Li Deng
3. Jinyu Li
4. Yoshua Bengio
5. Shinji Watanabe
6. Tara N. Sainath
7. Daniel Povey
8. Quoc V. Le

In Signal Processing

1. Dong Yu
2. Jinyu Li
3. Li Deng
4. Daniel Povey
5. Shinji Watanabe
6. Sanjeev Khudanpur
7. S. Renals
8. Yifan Gong

Citations in the corpus

• Impact of deep MLP architecture on different acoustic modeling techniques for under-resourced speech recognition (2013, background): “Index Terms— KL-HMM, Tandem, hybrid system, deep MLPs, under-resourced speech recognition”
• Cross-language knowledge transfer using multilingual deep neural network with shared hidden layers (2013, methodology): “We believe this is partly because we used much larger data sets than that used in [14] for both the source and target languages.”
• Multilingual acoustic models using distributed deep neural networks (2013, background): “Transfer learning [27] refers to the situation where the network of a target language is initialized with an existing source network, for example, by pre-training or fine-tuning [21, 22, 19, 23], and is not trained from scratch.”
• Multilingual training of deep neural networks (2013, methodology): “This hypothesis is further supported by our earlier work [13], in which we showed that layerwise pretraining of deep networks using stacked restricted Boltz-mann machines (RBMs) [14] is not sensitive to the choice of language.”
• Elastic spectral distortion for low resource speech recognition with deep neural networks (2013, background): “Those works focused on how to transfer the acoustics of other resource-rich languages to the target language [7, 8] or to make the language-independent feature front-end [9].”
• Hybrid acoustic models for distant and multichannel large vocabulary speech recognition (2013, methodology): “The various hyper-parameters are same as those described in [30], except the initial learning rate, which was tuned to 0.”
• Investigation of transfer learning for ASR using LF-MMI trained neural networks (2017, methodology): “Unsupervised pretraining has also been applied to multilingual speech recognition [9].”
• Semi-supervised training of Deep Neural Networks (2013, result): “However, there is no WER difference between the two systems, this is consistent with observations previously published in [25][26].”
• Improving low-resource CD-DNN-HMM using dropout and multilingual DNN training (2013, background): “There has been a number of works dedicated to training multilingual bottleneck features [11, 12, 13] or probabilistic posterior features [5, 12] used in tandem systems.”
• An open-source end-to-end ASR system for Brazilian Portuguese using DNNs built from newly assembled corpora (2020, background): “[12] did an unsupervised pretraining on different languages while Dalmia et al.”
• Transfer learning using computational intelligence: A survey (2015, methodology): “[52] exploited untranscribed acoustic data to the target languages in a deep neural network on unsupervised cross-lingual knowledge transfer.”
• Deep maxout networks for low-resource speech recognition (2013, background): “Multilingual network training enables crosslanguage knowledge transfer and can happen either in pretraining [13] or in the fine-tuning stage [11, 12].”
• Multilingual deep neural network based acoustic modeling for rapid language adaptation (2014, background): “In (deep) Tandem systems [5] on the other hand, each HMM state considers the whole MLP output vector.”
• Sequence-Based Multi-Lingual Low Resource Speech Recognition (2018, background): “Authors in [12] looked at unsupervised pretraining on different languages for a cross lingual recognition.”
• Multilingual representations for low resource speech recognition and keyword search (2015, background): “Multilingual (ML) models have been shown to outperform unilingual models for ASR in low resource languages [1–8].”