Weak top-down constraints for unsupervised acoustic model training

By Aren Jansen, Samuel Thomas, Hynek Hermansky. 2013

61 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

Unsupervised acoustic unit discovery for speech synthesis using discrete latent-variable neural networks (2019, background): “A central problem is finding frame-level feature representations that capture meaningful linguistic contrasts, such as phonetic categories, while being invariant to factors such as a speaker’s identity, gender or accent [2–5].”
A segmental framework for fully-unsupervised large-vocabulary speech recognition (2016, methodology): “The approach is further motivated by the observation that it is often easier to identify cross-speaker similarities between words than be-tween subwords [32], which is why most UTD systems focus on longer-spanning patterns.”
Unsupervised Learning of Semantic Audio Representations (2017, background): “archy: repeated unknown words have the same unknown phonetic structure [15], conversations with same unknown topic have shared unknown words [16], etc.”
A summary of the 2012 JHU CLSP workshop on zero resource speech technologies and models of early language acquisition (2013, methodology): “Likewise, stripped of any guidance from word transcripts and a pronunciation dictionary, the normal expectation-maximization training procedures for Gaussian mixturebased acoustic models are no longer capable of identifying speaker independent phonetic categories in a purely bottom-up fashion [24].”
Discriminative acoustic word embeddings: Tecurrent neural network-based approaches (2016, unknown): “There have now been a number of approaches compared on this same task and data [12, 21, 22, 23].”
Multilingual bottle-neck feature learning from untranscribed speech (2017, background): “Posteriorgrams can be derived from a Gaussian mixture model (GMM) [4, 8], sub-clustered GMM [9], unsupervised hidden Markov models (HMMs) [10, 11], or deep Boltzmann machine (DBM) [12].”
Towards Visually Grounded Sub-word Speech Unit Discovery (2019, background): “Some proposed approaches operate at the word level [1, 2, 3], while a separate line of work is concerned with sub-word modeling [4, 5, 6, 7].”
A Brief Overview of Unsupervised Neural Speech Representation Learning (2022, background): “Non-neural LVMs for speech representation learning have also been explored (Lee and Glass 2012; Ondel, Burget, and Cernocký 2016; Heck, Sakti, and Nakamura 2017; Jansen, Thomas, and Hermansky 2013).”
ABX-Discriminability Measures and Applications. (Mesures de discriminabilité ABX et applications) (2016, methodology): “6% AP on Switchboard for the GMM-based method from [87] vs.”
Spoken Content Retrieval—Beyond Cascading Speech Recognition with Text Retrieval (2015, methodology): “2) Top-down Constraints: It has been observed that wordlevel patterns are easier to identify across speakers than phoneme-level ones [241].”
Unsupervised neural network based feature extraction using weak top-down constraints (2015, methodology): “For all three hidden unit settings, performance is within 12% relative to the respective optimal settings for networks with 7 to 21 layers.”
Unsupervised Word Segmentation and Lexicon Discovery Using Acoustic Word Embeddings (2016, background): “between subwords [25], which is why most UTD systems focus on longer-spanning patterns.”
Query-by-Example Search with Discriminative Neural Acoustic Word Embeddings (2017, result): “The size of this set is comparable to those used for training in prior work on acoustic word embeddings [13, 27, 28].”
Deep convolutional acoustic word embeddings using word-pair side information (2015, methodology): “For testing, we use the 11k-token set Ytest from [3, 10, 11], making the results from”
Virtual Phone Discovery for Speech Synthesis Without Text (2019, methodology): “Gaussian universal background model posteriorgrams were matched using dynamic time warping to improve spoken term discovery (STD) [8].”