fMPE: discriminatively trained features for speech recognition

By Daniel Povey, Brian Kingsbury, Lidia Mangu, George Saon, Hagen Soltau, Geoffry Zweig. Proceedings. (ICASSP '05). IEEE International Conference on Acoustics, Speech, and Signal Processing, 2005., 2005

331 citations recorded by the source corpus.

Subfields: Artificial Intelligence

View Paper: https://www.semanticscholar.org/paper/1a70ac6680d319905d6bfca4cea0b4dc6c15f420

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. Li Deng
2. Dong Yu
3. Jinyu Li
4. Yoshua Bengio
5. S. Renals
6. James R. Glass
7. Jianfeng Gao
8. Tara N. Sainath

Citations in the corpus

• Hierarchical learning : theory with applications in speech and vision (2009, background): “Povey and collaborators [81] have however described a framework for estimating robust features from distorted cepstral coefficients that are typically added to the standard features, while other authors have proposed variations on standard MFCCs that provide additional robustness in mismatched condi”
• Multi-level adaptive networks in tandem and hybrid ASR systems (2013, methodology): “However, HMM-GMM systems became increasingly more accurate owing to the use various techniques such as contextdependent phone modelling [8] speaker adaptation using MLLR [9], sequence-level discriminative training techniques such as MMI and MPE [10], and high-dimension feature space transforms such ”
• Data augmentation for low resource languages (2014, methodology): “In addition to SI, the third stage also performs CMLLR-based speaker adaptive training (SAT) first using ML [21] and then feature-space MPE (fMPE-SAT) criterion [42].”
• Speech Recognition Using Augmented Conditional Random Fields (2009, background): “In some cases, it is computationally more efficient to focus on dense regions in the observation space and construct spaces with dimensionality .”
• Analysis of MLP-Based Hierarchical Phoneme Posterior Probability Estimator (2011, background): “G8: Frame-Based MPE: The hierarchical system discussed in this work can be related to the frame-based minimum phone error (fMPE) system [36].”
• Investigation of unsupervised adaptation of DNN acoustic models with filter bank input (2014, methodology): “Furthermore, existing discriminative feature mapping techniques, such as fMPE [21], may be applied on top, although this is not investigated in this work.”
• Dynamic Speech Models: Theory, Algorithms, and Applications (2006, methodology): “Over the past decade or so, there has been a popular discriminant-function-based and conditional modeling approach to speech recognition, making use of HMMs (as a discriminant function instead of as a generative model) or otherwise [13, 15–19].”
• Model-Based Approaches to Handling Uncertainty (2011, methodology): “Due to this reason, and the added problem of delta and delta-delta parameters, featureenhancement schemes based on stereo training [3] are used to combine noise robustness with state-of-the-art front-end processing such as semi-tied transforms [19] and fMPE [51].”
• Deep and Wide: Multiple Layers in Automatic Speech Recognition (2012, methodology): “At IBM, researchers developed a technique called Featurebased Minimum Phone Error (fMPE) [16], which incorporated the MPE error criterion at the feature level.”
• Extracting deep neural network bottleneck features using low-rank matrix factorization (2014, methodology): “Such improvements include better use of context information via fMPE [22], and using speaker-adapted features as DNN inputs.”
• Speech recognition and keyword spotting for low-resource languages: Babel project research at CUED (2014, methodology): “A speaker adaptive training (SAT) system using global constrained maximum likelihood linear regression (CMLLR) at a speaker level [35], was then constructed incorporating both Minimum Phone Error (MPE) [36] training and feature-space MPE (fMPE) [37].”
• Semi-supervised GMM and DNN acoustic model training with multi-system combination and confidence re-calibration (2013, methodology): “The 100 hour voice search data with six versions of transcription including the human transcription are used to train MLE (maximum likelihood estimation), fMPE (feature-minimum phone error) [13], and fMPE+bMMI (boosted maximum mutual information) [14] models separately.”
• Pushing the envelope - aside [speech recognition] (2005, methodology): “Finally, we note the related and impressive work done recently at IBM on incorporating low-dimensional projections of high-dimensional, Gaussian-derived posteriors, also incorporated in the feature vector used for recognition [ 25 ].”
• Large-margin minimum classification error training: A theoretical risk minimization perspective (2008, background): “Discriminative training for hidden Markov models (HMMs) has become one of the core technologies in speech recognition research over the past dozen years or so (Juang et al., 1997; Juang and Katagiri, 1992; McDermott, 1997; McDermott et al., 2007; Povey et al., 2004; Woodland and Povey, 2000).”
• The IBM 2015 English conversational telephone speech recognition system (2015, methodology): “2% WER during the 2004 DARPA EARS Rich Transcription evaluation [5] largely due to the Attila ASR toolkit [6] and fMPE [7].”