Multilingual MLP features for low-resource LVCSR systems

By Samuel Thomas, Sriram Ganapathy, Hynek Hermansky. 2012

122 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. Jinyu Li
2. Dong Yu
3. Li Deng
4. Tara N. Sainath
5. James R. Glass
6. Shinji Watanabe
7. Yoshua Bengio
8. Quoc V. Le

In Speech & Audio

1. Jinyu Li
2. Dong Yu
3. Tara N. Sainath
4. Shinji Watanabe
5. James R. Glass
6. Yu Zhang
7. Wei-Ning Hsu
8. Karen Livescu

In Signal Processing

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

Citations in the corpus

• Multilingual Speech Recognition with a Single End-to-End Model (2017, background): “…continuous speech recognizers are severly limited [1] These challenges have meant that there has been a perennial interest in multilingual and cross-lingual models which allow for knowledge transfer across languages, and thus relieve burdensome data requirements [2, 3, 4, 5, 6, 7, 8, 9, 10, 11].”
• Pre-training on high-resource speech recognition improves low-resource speech-to-text translation (2018, background): “On the speech side, it might be even more effective to use multilingual training; or to replace the MFCC input features with pre-trained multilingual features, or features that are targeted to low-resource multispeaker settings (Kamper et al., 2015, 2017; Thomas et al., 2012; Cui et al., 2015; Yuan ”
• Massively Multilingual Adversarial Speech Recognition (2019, background): “…trained on one language to effectively recognize the sounds of other languages (Schultz and Waibel, 2001; Le and Besacier, 2005; Stolcke et al., 2006; T´oth et al., 2008; Plahl et al., 2011; Thomas et al., 2012; Imseng et al., 2014; Do et al., 2014; Heigold et al., 2013; Scharenborg et al., 2017).”
• 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.”
• Large-Scale Multilingual Speech Recognition with a Streaming End-to-End Model (2019, background): “Successful strategies for building multilingual acoustic models (AMs) include stacked bottleneck features [1–4], shared hidden layers [5,6], knowledge distillation [7], and multitask learning [8].”
• The language-independent bottleneck features (2012, background): “For ANNs, adaptation to new language is possible [9][10].”
• Multi-level adaptive networks in tandem and hybrid ASR systems (2013, methodology): “Aside from Thomas et al [24], other schemes operate only on the outputs of first-level MLPs, without the original features, limiting the power of the second-level DNN; or they use shallow MLPs [30], so that the final structure is effectively one single DNN without pre-training.”
• Cross-language knowledge transfer using multilingual deep neural network with shared hidden layers (2013, methodology): “Another difference between ours and that in [24] is that we used senones as the DNN outputs and used DNNs instead of shallow MLPs.”
• Automatic speech recognition for under-resourced languages: A survey (2014, background): “[Thomas, 2012], [Thomas 2012b], and [Vesely, 2012] demonstrated how to use data from multiple languages to extract features for an under-resourced language and, hence, improve ASR performance.”
• Investigation on cross- and multilingual MLP features under matched and mismatched acoustical conditions (2013, methodology): “Compared to [14], instead of using the data of different languages sequentially, our BN are trained on the merged multilingual resources jointly.”
• An Investigation of Deep Neural Networks for Multilingual Speech Recognition Training and Adaptation (2017, methodology): “Thus, bootstrapping a language with little or no data from another language by using a tandem approach [2, 4, 17, 18, 19], initialization from an existing network [20, 21] or model adaptation [22] is beyond the scope of this paper.”
• Deep neural network features and semi-supervised training for low resource speech recognition (2013, methodology): “While this work is related to several recent approaches [8, 9, 10, 11, 12], we use two different techniques to derive better features and improve acoustic model training in low resource settings: data driven features extracted using deep neural networks (DNN) This work was supported in parts by the…”
• Multilingual training of deep neural networks (2013, methodology): “Ofthese, the method of Thomas, et al. [20] is the closest in spirit to our current work.”
• Unsupervised cross-lingual knowledge transfer in DNN-based LVCSR (2012, methodology): “A number of techniques for cross-lingual acoustic modelling have been published including the use of: global phone sets [1, 2]; multilingual posterior features in tandem [3, 4] and Kullback-Liebler hidden Markov model (KL-HMM) systems [5]; subspace Gaussian mixture models (SGMMs) with a shared multi”