A compact model for speaker-adaptive training

By Tasos Anastasakos, J. McDonough, Richard M. Schwartz, John I. Makhoul. 1996

We formulate a novel approach to estimating the parameters of continuous density HMMs for speaker-independent (SI) continuous speech recognition. It is motivated by the fact that variability in SI acoustic models is attributed to both phonetic variation and variation among the speakers of the training population, that is independent of the information content of the speech signal. These two variation sources are decoupled and the proposed method jointly annihilates the inter-speaker variation and estimates the HMM parameters of the SI acoustic models. We compare the proposed training algorithm to the common SI training paradigm within the context of supervised adaptation. We show that the proposed acoustic models are more efficiently adapted to the test speakers, thus achieving significant overall word error rate reductions of 19% and 25% for 20K and 5K vocabulary tasks respectively.

Subfields: Speech & Audio

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Speech & Audio

Citations in the corpus

Librispeech: An ASR corpus based on public domain audio books (2015, methodology): “In this second decoding stage, we use a speaker-adapted model [18, 19] with fMLLR transforms estimated at the speake r level, based on the transcript generated by the first decodin g pass.”
Speaker Representations for Speaker Adaptation in Multiple Speakers' BLSTM-RNN-Based Speech Synthesis (2016, background): “Speaker adaptive training is one of the most effective approaches to train a high quality acoustic model with a limited database [5, 6, 9, 10].”
Improving Robustness of Deep Neural Network Acoustic Models via Speech Separation and Joint Adaptive Training (2015, methodology): “ical models performs better than transforming the models trained in the original input space [2], [15].”
A scalable approach to using DNN-derived features in GMM-HMM based acoustic modeling for LVCSR (2013, background): “, [16,20]) or irrelevant variability normalization (IVN) based training (e.”
Rapid speaker adaptation in eigenvoice space (2000, methodology): “The reference models are obtained by an iterative scheme called " cluster adaptive training " (CAT) which is related to speaker-adaptive training (SAT) [4], [5].”
Unsupervised Adaptation With Discriminative Mapping Transforms (2009, methodology): “As unsupervised adaptation is the focus of this work, the 1-best hypothesis generated from the MLLR adapted model was used as the reference transcription.”
Analysis of Speaker Adaptation Algorithms for HMM-Based Speech Synthesis and a Constrained SMAPLR Adaptation Algorithm (2009, methodology): “Because this training data includes a lot of speaker-dependent characteristics that affect the adapted models and the quality of synthetic speech generated from them, we incorporated the speaker-adaptive training (SAT) algorithm [33] into our speech synthesis system in order to reduce the negative i”
Embedding-Based Speaker Adaptive Training of Deep Neural Networks (2017, methodology): “Speaker adaptation [2][3] and speaker adaptive training (SAT) [4][5] have been a common practice in ASR to reduce the covariate shift incurred by speaker variability.”
Machine Learning Paradigms for Speech Recognition: An Overview (2013, unknown): “been developed, unique in ASR and not seen in the ML literature, in the frameworks of speaker adaptive training (SAT) [205] and noise adaptive training (NAT) [201], [206], [207].”
Learning Spectral Mapping for Speech Dereverberation and Denoising (2015, methodology): “To train the neural network, the input is normalized to zero mean and unity variance over all feature vectors in the training set, and the output is normalized into the range of [0] [1].”
Feature optimized DPGMM clustering for unsupervised subword modeling: A contribution to zerospeech 2017 (2017, background): “• Maximum likelihood linear transforms (MLLT) to reduce complexity and to de-correlate features [13, 14], • Feature-space maximum likelihood linear regression (fMLLR) to reduce speaker variability within speech features [15, 16],”
Speaker Adaptive Training of Deep Neural Network Acoustic Models Using I-Vectors (2015, background): “Another technique closely related with speaker adaptation is speaker adaptive training (SAT) [14], [15].”
Speech Synthesis Based on Hidden Markov Models (2013, methodology): “The average voice model is a “canonical” speakerindependent HMM where inter-speaker acoustic variation is normalized using a technique called on speaker-adaptive training (SAT) [60], [61].”
Statistical Parametric Speech Synthesis (2007, background): “They also reported that “recording condition-adaptive training,” which is based on the same idea as speaker-adaptive training (Anastasakos et al., 1996; Gales, 1998), worked effectively to normalize recording conditions.”
An Overview of Noise-Robust Automatic Speech Recognition (2014, background): “All of these model-space joint model training methods share the same spirit with speaker adaptive training (SAT) [278], proposed in 1996 for speaker adaptation.”
Robust Speaker-Adaptive HMM-Based Text-to-Speech Synthesis (2009, methodology): “Previously, we had utilized a model-space SAT algorithm [38] using linear transformations of mean vectors of Gaussian pdfs in our average voice systems [32], [39].”
Large-vocabulary speech recognition under adverse acoustic environments (2000, methodology): “This is called "Noisy Adaptive Training" (NAT) and is an extension of the earlier "Speaker Adaptive Training" [1] to noise-robustness.”