Speaker Adaptive Training of Deep Neural Network Acoustic Models Using I-Vectors

By Yajie Miao, Hao Zhang, Florian Metze. 2015

124 citations recorded by the source corpus.

Subfields: Artificial Intelligence, 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. Yoshua Bengio
5. Florian Metze
6. Yifan Gong
7. Wei Ping
8. Shinji Watanabe

In Signal Processing

1. Jinyu Li
2. Dong Yu
3. Yoshua Bengio
4. Florian Metze
5. Yifan Gong
6. Li Deng
7. Wei Ping
8. Yajie Miao

Citations in the corpus

• Salvaging Federated Learning by Local Adaptation (2020, background): “Many techniques are not compatible with federated learning because they require an ensemble of models (Tan et al., 2015) or are speech-specific (Miao et al., 2015), but (Huang et al., 2015) connects personalization and multi-task learning.”
• Neural Voice Cloning with a Few Samples (2018, methodology): “Extraction of the embeddings can be based on i-vectors [Miao et al., 2015], or bottleneck layers of neural networks trained with a classification loss [Li and Wu, 2015].”
• EESEN: End-to-end speech recognition using deep RNN models and WFST-based decoding (2015, methodology): “We will study new speaker adaptation [35, 36] and adaptive training [37, 38] techniques for the CTC models.”
• Speaker-Invariant Training Via Adversarial Learning (2018, background): “Therefore, speaker adaptation meth-ods are widely used to boost the recognition system performance [3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13].”
• Embedding-Based Speaker Adaptive Training of Deep Neural Networks (2017, background): “In [17], i-vectors are mapped to speaker-dependent biases which are added to input features to compensate the speaker variability.”
• Dynamic Layer Normalization for Adaptive Neural Acoustic Modeling in Speech Recognition (2017, background): “Model-based adaptation typically brings more improvement than auxiliary feature-based adaptation as shown in [12].”
• Unspeech: Unsupervised Speech Context Embeddings (2018, background): “Several forms of context/speaker embeddings have also been used for (speaker) adaption in state-of-the-art speech recognition acoustic models: i-vector speaker embeddings are by far the most popular [1, 19, 20].”
• Unsupervised Adaptation of Acoustic Models for ASR Using Utterance-Level Embeddings from Squeeze and Excitation Networks (2019, background): “Many studies show that when i-vectors were appended with acoustic features in DNN training, it gives a consistent gain in various ASR tasks [19], [20], [21].”
• Factorized Hidden Layer Adaptation for Deep Neural Network Based Acoustic Modeling (2016, background): “Furthermore, it is possible to obtain speaker normalized features using an adaptation network [57].”
• A survey of deep neural network architectures and their applications (2017, methodology): “Moreover, the DNNs have been employed for tracking dialog state [53], transferring cros s-language knowledge [71], learning the filter banks [135], designing an automatic feature extraction systems from aud io [51], and speaker adaptive training (SAT) for acoustic modeling [114].”
• Speaker adaptive training in deep neural networks using speaker dependent bottleneck features (2016, methodology): “The proposed framework also has resemblance with the architecture proposed in [17], where I-vectors have been transformed to linear shifts using a DNN to perform speaker normalisation on the input features before training the SAT-DNN.”