Speaker adaptation of neural network acoustic models using i-vectors

By George Saon, Hagen Soltau, David Nahamoo, Michael A. Picheny. 2013

679 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

In Speech & Audio

In Signal Processing

Citations in the corpus

ContextNet: Improving Convolutional Neural Networks for Automatic Speech Recognition with Global Context (2020, background): “Previous works on hybrid ASR have successfully introduced the context to acoustic models by either stacking a large number of layers, or having a separately trained global vector to represent the speaker and the environment information [13, 14, 15, 16].”
Wav2Letter: an End-to-End ConvNet-based Speech Recognition System (2016, methodology): “xtend this training pipeline. 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]. Recently, [23] proposed GMM-free training, but the approach still requires to generate a f”
Personalization of End-to-End Speech Recognition on Mobile Devices for Named Entities (2019, background): “There are existing speaker adaptation techniques for neural networks that offer a more compact model representation by having a large number of global parameters shared across different speakers and a smaller number of speaker-specific parameters [13, 14, 15, 16, 17, 18, 19, 20].”
Deep Learning: Methods and Applications (2014, neutral): “, 2013b), IBM (Sainath et al., 2011, 2013, 2013b; Kingsbury et al., 2012; Saon et al., 2013), Google (Jaitly et al.”
Achieving Human Parity in Conversational Speech Recognition (2016, methodology): “Speaker adaptive modeling in our system is based on conditioning the network on an i-vector [49] characterization of each speaker [50, 51].”
Letter-Based Speech Recognition with Gated ConvNets (2017, background): “The current state of the art on LibriSpeech belongs to this approach too (Panayotov et al., 2015; Peddinti et al., 2015b), with an additional step of speaker adaptation (Saon et al., 2013; Peddinti et al., 2015a).”
The microsoft 2016 conversational speech recognition system (2016, methodology): “Compared to the networks used previously in image recognition, this network uses small (3x3) filters, is deeper, and applies up to five convolutional layers before pooling.”
The CAPIO 2017 Conversational Speech Recognition System (2017, background): “, i-vectors [11] in case of speaker adaptation) for a better feature transformation [12, 13, 14].”
Extracting Domain Invariant Features by Unsupervised Learning for Robust Automatic Speech Recognition (2018, methodology): “We apply the same procedure as i-vector based speaker adaptation [27]: For each utterance, we first estimate its s-vector, and then concatenate s-vectors with the feature representation of each frame to generate the new feature sequence.”
Exploring Machine Speech Chain for Domain Adaptation and Few-Shot Speaker Adaptation (2021, methodology): “In ASR, speaker adaptation techniques can be divided into three categories [11], namely model retraining [12, 13], feature transformation [14, 15] and auxiliary feature-based adaptation [16, 11].”
Joint training of convolutional and non-convolutional neural networks (2014, methodology): “The I-Vectors were generated exactly the same way as described in [18].”
A study on data augmentation of reverberant speech for robust speech recognition (2017, methodology): “We use 40-dimensional MFCCs and 100 dimensional i-vectors [21].”
Invariant Representations for Noisy Speech Recognition (2016, methodology): “Several model and feature based adaptation methods such as Maximum Likelihood Linear Regression (MLLR), feature-based MLLR and iVectors (Saon et al. , 2013) have been proposed to handle speaker variability; and Noise Adaptive Training (NAT; Kalinli et al. , 2010) and Vector Taylor Series (VTS; Un…”
Factorized Hidden Layer Adaptation for Deep Neural Network Based Acoustic Modeling (2016, background): “Techniques like i-vectors [12], [13], [61], [62] and bottleneck features [63]–[65] are commonly used as speaker features.”
The Microsoft 2017 Conversational Speech Recognition System (2017, methodology): “Speaker adaptive modeling in our system is based on conditioning the network on an i-vector [37] characterization of each speaker [38, 39].”
Recent progresses in deep learning based acoustic models (2017, methodology): “One popular method in this category is the use of auxiliary features, such as i-vector [106], [107], speaker code [108] , and noise estimate [109] which are concatenated with the standard acoustic features.”
Hierarchical Bayesian combination of plug-in maximum a posteriori decoders in deep neural networks-based speech recognition and speaker adaptation (2017, methodology): “…as testified by the great deal of DNN adaptation techniques available in the recent literature, e.g., (Yu et al., 2013; AbdelHamid and Jiang, 2013; Saon et al., 2013; Siniscalchi et al., 2013; Li et al., 2014; Remes et al., 2015; Huang et al., 2015a; Swietojanski et al., 2016; Huang et al., 2017).”