Large scale deep neural network acoustic modeling with semi-supervised training data for YouTube video transcription

By Hank Liao, Erik McDermott, Andrew W. Senior. 2013

216 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. Yu Zhang
4. Tara N. Sainath
5. Yoshua Bengio
6. Quoc V. Le
7. Li Deng
8. Yonghui Wu

In Speech & Audio

1. Jinyu Li
2. Yu Zhang
3. Tara N. Sainath
4. Yonghui Wu
5. Dong Yu
6. Gabriel Synnaeve
7. Wei-Ning Hsu
8. Chung-Cheng Chiu

In Signal Processing

1. Jinyu Li
2. Dong Yu
3. Tara N. Sainath
4. Yu Zhang
5. Shinji Watanabe
6. Yonghui Wu
7. Li Deng
8. Bo Li

Citations in the corpus

• Robust Speech Recognition via Large-Scale Weak Supervision (2022, background): “Liao et al. (2013) is an early example of leveraging weakly supervised learning to increase the size of a deep learning based speech recognition dataset by over 1,000 hours.”
• BigSSL: Exploring the Frontier of Large-Scale Semi-Supervised Learning for Automatic Speech Recognition (2021, methodology): “Besides the obvious advantage of having a very large amount of audio available, these YouTube-based datasets have an extremely wide range of sub-domains [54], [69], plotted in Fig.”
• Toward Domain-Invariant Speech Recognition via Large Scale Training (2018, methodology): “of the sets used for training and evaluation are anonymized, and are representative of Google’s voice search, captioning and cloud traffic. The YouTube set was transcribed in a semi-supervised fashion [28, 29]; the rest of the datasets are hand-transcribed. 4.2. Dataset analysis To understand the dat”
• Iterative Pseudo-Labeling for Speech Recognition (2020, neutral): “Semi-supervision in ASR is well-studied [15, 16, 17, 18, 19, 20]; our work builds primarily on recent work with end-to-end systems, especially PL [9, 1].”
• ContextNet: Improving Convolutional Neural Networks for Automatic Speech Recognition with Global Context (2020, methodology): “We use a experiment setup similar to [35], where the training set has public Youtube videos with semi-supervised transcripts generated by the approach in [36].”
• Gender and Dialect Bias in YouTube’s Automatic Captions (2017, background): “The first is that they’re backed by Google’s speech recognition software, which is both very popular and among the more accurate proprietary ASR systems (Liao et al., 2013).”
• Deep Learning: Methods and Applications (2014, background): “vocabulary tasks with more convincing, highly positive results (Yu et al., 2010c; Dahl et al., 2011, 2012; Seide et al., 2011; Kubo et al., 2012; Hinton et al., 2012; Kingsbury et al., 2012; Deng et al., 2013a, 2013b; Su et al., 2013; Yan et al., 2013; Liao et al., 2013).”
• LibriTTS: A Corpus Derived from LibriSpeech for Text-to-Speech (2019, methodology): “Auto-sync is also used to generate data for ASR acoustic model training [40].”
• Compression of fully-connected layer in neural network by Kronecker product (2015, background): “In Sainath et al. (2013), Liao et al. (2013), Xue et al. (2013), Zhang et al. (2014b), Denton et al. (2014), low rank approximation of fully-connected layer is used; and Jaderberg et al. (2014), Rigamonti et al. (2013), Lebedev et al. (2014) considered low rank approximation of convolution layer.”
• Lessons from Building Acoustic Models with a Million Hours of Speech (2019, methodology): “Several methods to estimate word and frame level token confidence from speech lattices or hypotheses have been developed ([21], [22]).”
• Building DNN acoustic models for large vocabulary speech recognition (2014, background): “Previous work explores factoring the output layer to increase the relative number of shared parameters [64], [65], but this effect occurs naturally by substantially increasing network size.”
• German End-to-end Speech Recognition based on DeepSpeech (2019, background): “It is a key component of voice assistants like Google Home (Li et al., 2017), in spoken language translation devices (Krstovski et al., 2008), or for automatic transcription of audio and video files (Liao et al., 2013).”
• Domain Adaptation Using Factorized Hidden Layer for Robust Automatic Speech Recognition (2018, background): “YouTube video captioning [8], on the other hand, deals with much longer speech data with different recording quality depending on the genre.”
• Large scale weakly and semi-supervised learning for low-resource video ASR (2020, background): “Reducing the amount of manual transcription for building speech recognition systems has been a constant research theme over the past three decades [40, 20, 9, 5, 41, 42] given the revolutionary increase and access to unlabeled data and computational resources.”
• Contrastive Semi-Supervised Learning for ASR (2021, methodology): “One important ingredient of pseudo-labeling is filtering the teacher generated data by removing implausible, low quality hypotheses through confidence filtering [14].”
• Identifying the sentiment styles of YouTube’s vloggers (2018, background): “…(LeCun et al., 2015; Zhang et al., 2017b; Zhang et al., 2017a) and contributed to an accuracy increase of fifty percent for YouTube’s automatic captions in English (Official YouTube Blog, 2017), the automatic generation of YouTube video captions remains a challenging task (Liao et al., 2013).”
• Bridging the gap between streaming and non-streaming ASR systems bydistilling ensembles of CTC and RNN-T models (2021, methodology): “As in [1], we train our teacher models on Confisland data [25].”
• Lattice rescoring strategies for long short term memory language models in speech recognition (2017, methodology): “It contained 13 categories, ranging from Science&Education to Video Games [30, 33, 31].”