A time delay neural network architecture for efficient modeling of long temporal contexts

By Vijayaditya Peddinti, Daniel Povey, Sanjeev Khudanpur. 2015

1,083 citations recorded by the source corpus.

Subfields: Artificial Intelligence, Model Architectures, Efficiency & Compression, 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 Model Architectures

In Efficiency & Compression

In Signal Processing

Citations in the corpus

TED-LIUM 3: twice as much data and corpus repartition for experiments on speaker adaptation (2018, methodology): “We present experimental results with different ASR architectures, by using Time Delay Neural Network (TDNN) [5] and Factored TDNN (TDNN-F) acoustic models [7] on the legacy distribution of TED-LIUM 3 in section 3, and also exploring the use of a pure neural endto-end system in section 4.”
Wav2Letter: an End-to-End ConvNet-based Speech Recognition System (2016, methodology): “The current state of the art on Librispeech (the dataset that we used for our evaluations) uses this approach too [18, 20], with an additional step of speaker adaptation [22, 19].”
Federated Learning for Keyword Spotting (2018, methodology): “The architecture is a CNN with 5 stacked dilated convolutional layers of increasing dilation rate, followed by two fully-connected layers and a softmax inspired from [15].”
Letter-Based Speech Recognition with Gated ConvNets (2017, methodology): “Current state of the art models on LibriSpeech also employ this approach (Panayotov et al., 2015; Peddinti et al., 2015b), with an additional step of speaker adaptation (Saon et al., 2013; Peddinti et al., 2015a).”
Transformer-Based Acoustic Modeling for Hybrid Speech Recognition (2019, background): “To address these issues, a variety of neural network architectures have been proposed to replace RNNs, including time delay neural networks (TDNN) [5], feed-forward sequential memory networks (FSMN) [6], and convolution neural networks (CNN) [4,13], while only limited success has been achieved.”
Snips Voice Platform: an embedded Spoken Language Understanding system for private-by-design voice interfaces (2018, methodology): “We train a deep neural network, consisting of time-delay layers similar to those presented in [39], and long short-term memory layers similar to those of [40].”
AISHELL-2: Transforming Mandarin ASR Research Into Industrial Scale (2018, methodology): “Based on the tied-triphone state alignments from GMM, a time-delayed neural network (TDNN, [12]) is then configured and trained.”
Single-Channel Multi-talker Speech Recognition with Permutation Invariant Training (2017, background): “Thanks to the signiﬁcant progresses made in the recent years [1], [2], [3], [4], [5], [6], [7], [8], [9], [10], [11], [12], [13], [14], [15], [16], [17], [18], [19], [20], the ASR systems now surpassed the threshold for adoption in many real-world scenarios and enabled services such as Microsoft…”
AISHELL-1: An open-source Mandarin speech corpus and a speech recognition baseline (2017, methodology): “The DNN model we used was the time delay neural network (TDNN) [9].”
Improving Speech Recognition by Revising Gated Recurrent Units (2017, background): “The early deep learning works in speech recognition were mainly based on standard multilayer perceptrons (MLPs) [6, 7], while recent systems benefit from more advanced architectures, such as Convolutional Neural Networks (CNNs) [8] and Time Delay Neural Network (TDNN) [9, 10].”
Recent progresses in deep learning based acoustic models (2017, background): “…which convolve along both the frequency and time axes and thus exploit variable-length contextual information, attracted new attentions, this time under the DL/HMM hybrid architecture [13], [49] and with variations such as row convolution [15] and feedforward sequential memory network (FSMN) [16].”
SpeechMoE: Scaling to Large Acoustic Models with Dynamic Routing Mixture of Experts (2021, methodology): “Various types of neural network architectures have been employed in ASR systems, such as convolutional neural networks (CNNs) [3, 4], long short-term memory (LSTM) [5], gated recurrent unit[6], time-delayed neural network [7], feedforward sequential memory networks (FSMN) [8], etc.”
Investigating Generative Adversarial Networks based Speech Dereverberation for Robust Speech Recognition (2018, methodology): “We used Kaldi to train our backend ASR system with the similar acoustic model architecture and features in [28].”
A novel pyramidal-FSMN architecture with lattice-free MMI for speech recognition (2018, methodology): “To achieve better results, we train a 5-layer TDNN-LSTM LM for rescoring.”
Acoustic Modeling for Distant Multi-talker Speech Recognition with Single- and Multi-channel Branches (2019, background): “Table 1 shows the effect these data augmentation techniques had using the CHiME-5 baseline TDNN architecture.”
An empirical evaluation of zero resource acoustic unit discovery (2017, methodology): “40dimensional MFCCs (without cepstral truncation [18]) augmented with 3-dimensional pitch and probability of voicing features are used as inputs to the network.”
To Reverse the Gradient or Not: an Empirical Comparison of Adversarial and Multi-task Learning in Speech Recognition (2018, methodology): “1D ConvNets were introduced early in the speech community, and are also referred as Time-Delay Neural Networks [16].”
A Pruned Rnnlm Lattice-Rescoring Algorithm for Automatic Speech Recognition (2018, methodology): “All the acoustic models are trained using the open source speech recognition toolkit Kaldi, with either TDNN [13] or lattice-free maximum mutual information (MMI) [14] models; Backstitch optimization method [15] is used during acoustic model training.”