Connectionist Temporal Classiﬁcation: Labelling Unsegmented Sequences with Recurrent Neural Networks

By Alex Auvolat, Thomas Mesnard. 2016

Subfields: Natural Language Processing, Model Architectures

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Natural Language Processing

In Model Architectures

Citations in the corpus

Sequence to Sequence Learning with Neural Networks (2014, methodology): “The Connectionist Sequence Classiﬁcation is another popular technique for mapping sequences to sequences with neural networks, although it assumes a monotonic alignment between the inputs and the outputs [11].”
Deep Learning Scaling is Predictable, Empirically (2017, methodology): “We use Adam to optimize connectionist temporal classiﬁcation loss (CTC, Graves et al. (2006)).”
Mixed Precision Training (2017, methodology): “The model used for training on the English dataset consists of two 2D convolution layers, three recurrent layers with GRU cells, 1 row convolution layer and Connectionist temporal classiﬁcation (CTC) cost layer (Graves et al., 2006).”
Imputer: Sequence Modelling via Imputation and Dynamic Programming (2020, methodology): “Conveniently, the log term in Equation 12 can be computed exactly and efﬁciently via dynamic programming (Graves et al., 2006).”
Multimodal Transformer for Unaligned Multimodal Language Sequences (2019, methodology): “The CTC (Graves et al., 2006) module we use comprises two components: alignment predictor and the CTC loss.”
mSLAM: Massively multilingual joint pre-training for speech and text (2022, methodology): “…text from 101 languages. mSLAM is a multilingual extension of SLAM (Bapna et al., 2021), with the addition of a Connectionist Temporal Classification (CTC) loss (Graves et al., 2006) on the paired speech-text data, to reduce interference and encourage stronger alignment across the two modalities.”
XLS-R: Self-supervised Cross-lingual Speech Representation Learning at Scale (2021, methodology): “Finally, we evaluate our approach on language identification for which we use our smallest model as the tasks requires less capacity given the lower complexity compared to the structured prediction problems of speech recognition and speech translation.”
BigSSL: Exploring the Frontier of Large-Scale Semi-Supervised Learning for Automatic Speech Recognition (2021, positive): “These features are either used as input to an RNN transducer [33] along with a 2-layer LSTM decoder, or used as input for a connectionist temporal classiﬁcation (CTC) model [34] after an additional projection layer.”
Unsupervised Speech Recognition (2021, background): “Such models are often classified into two categories: hybrid (Juang et al., 1986; Young, 1996; Povey, 2005; Hinton et al., 2012; Abdel-Hamid et al., 2012; Bourlard and Morgan, 2012) or end-to-end (Graves et al., 2006; Graves, 2012; Chorowski et al., 2015; Rao et al., 2017).”
Robust wav2vec 2.0: Analyzing Domain Shift in Self-Supervised Pre-Training (2021, methodology): “Pre-trained models are ﬁne-tuned with connetionist temporal classiﬁcation [40].”
MLS: A Large-Scale Multilingual Dataset for Speech Research (2020, methodology): “We chose ASG criterion over Connectionist Temporal Classiﬁca-tion [9] (CTC) criterion since ASG doesn’t exhibit delay in transcriptions compared to CTC [10].”
Self-Training and Pre-Training are Complementary for Speech Recognition (2020, positive): “After pre-training on the unlabeled data, this model is ﬁne-tuned on the labeled data using Connectionist Temporal Classiﬁcation (CTC; [35]) and a letter-based output vocabulary.”
Rethinking Evaluation in ASR: Are Our Models Robust Enough? (2020, methodology): “Acoustic Model (AM) All models are trained with Connectionist Temporal Classiﬁcation [Graves et al., 2006] and the network architecture follows [Synnaeve et al., 2019]: the encoder of our AMs is composed of a convolutional frontend (1-”
Common Voice: A Massively-Multilingual Speech Corpus (2019, neutral): “(CTC) loss function (Graves et al., 2006).”
TED-LIUM 3: twice as much data and corpus repartition for experiments on speaker adaptation (2018, methodology): “The system is trained end-to-end by using the CTC loss function [2], in order to predict a sequence of characters from the input audio.”
Deep Speech 2 : End-to-End Speech Recognition in English and Mandarin (2015, methodology): “After the bidirectional recurrent layers we apply one or more fully connected layers with The output layer L is a softmax computing a probability distribution over characters given by The model is trained using the CTC loss function [22].”