Improving DNN speaker independence with I-vector inputs

By Andrew W. Senior, Ignacio Lopez-Moreno. 2014

210 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

Toward Domain-Invariant Speech Recognition via Large Scale Training (2018, methodology): “We extract 32-dimensional i-vectors, which capture the acoustic characteristics of an utterance [30].”
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].”
Parallel training of Deep Neural Networks with Natural Gradient and Parameter Averaging (2014, methodology): “Some other authors [33] use log Mel ﬁlterbank energies; the MFCC features we use here are equivalent to log Mel ﬁlterbank energies because MFCCs are a linear transform of them (we use the same number of coeﬃcients as ﬁlterbanks, 40 for these experiments) and our input data normalization (Section…”
Teacher-Student Training for Text-Independent Speaker Recognition (2018, unknown): “adaptation [5]) and emotion recognition [6].”
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].”
Low-rank plus diagonal adaptation for deep neural networks (2016, background): “The speaker-specific auxiliary features can be derived in many different ways, such as i-vector [15, 16, 17] and speaker code [18, 19, 20], and then are concatenated with the standard acoustic feature vectors.”
On speaker adaptation of long short-term memory recurrent neural networks (2015, methodology): “Third, the acoustic model is trained (and decoded) using speaker-adaptive (SA) features [22, 23, 24] or features enriched with SD information [25, 26, 27, 28].”
Multi-basis adaptive neural network for rapid adaptation in speech recognition (2015, methodology): “, using i-vector [11, 12] or underlying factors [13] as additional input features.”
Domain Adaptation Using Factorized Hidden Layer for Robust Automatic Speech Recognition (2018, methodology): “Many DNN adaptation methods have been previously studied, including i-vector feature augmentation [19, 20], speaker-code adaptation [21], cluster adaptive training (CAT) [22, 23], factorized hidden layer [9], learning hidden unit contribution (LHUC) [24], singular value decomposition based subspace…”
To Reverse the Gradient or Not: an Empirical Comparison of Adversarial and Multi-task Learning in Speech Recognition (2018, methodology): “sanInternatFacebookAIResearch. *Equal contribution Traditionally, the identity of the speakers has been used to extract speaker representations and use it as an additional input to the acoustic model [1, 2]. This approach requires an additional feature extraction process of the input signal. Recentl”
Cluster adaptive training for deep neural network (2015, methodology): “The representation can be estimated either independent of the DNN, such as the iVector adaptation [14, 15, 16], or dependent on the DNN, such as speaker code adaptation [17, 18].”
An Investigation Into On-device Personalization of End-to-end Automatic Speech Recognition Models (2019, unknown): “A common solution to this problem is to adapt a well-trained model using a small amount of user-specific data to build a personalized model for each user [4, 5, 6, 7, 8, 9, 10, 11, 12, 13].”
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.”
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.”
Transfer learning for speech and language processing (2015, unknown): “(although it is possible) [88], [89], [90], [91].”