## Hidden-Articulator Markov Models For Speech Recognition (2000)

### Cached

### Download Links

Venue: | In Proc. IEEE Intl. Conf. on Acoustics, Speech, and Signal Processing |

Citations: | 85 - 20 self |

### BibTeX

@INPROCEEDINGS{Richardson00hidden-articulatormarkov,

author = {Matt Richardson and Jeff Bilmes and Chris Diorio},

title = {Hidden-Articulator Markov Models For Speech Recognition},

booktitle = {In Proc. IEEE Intl. Conf. on Acoustics, Speech, and Signal Processing},

year = {2000},

pages = {713--716}

}

### Years of Citing Articles

### OpenURL

### Abstract

In traditional speech recognition using Hidden Markov Models (HMMs), each state represents an acoustic portion of a phoneme. We explore the concept of an articulator based HMM, where each state represents a particular articulatory configuration [Erler 1996]. In this paper, we present a novel articulatory feature mapping and a new technique for model initialization. In addition, we use diphone modeling which allows context dependent training of transition probabilities. Our goal is to confirm that articulatory knowledge can assist speech recognition. We demonstrate this by showing that our mapping of articulatory configurations to phonemes performs better than random mappings. Furthermore, we demonstrate the practicality of the model by showing that, in combination with a standard model, a 12-21% relative word error rate decrease occurs relative to the standard model alone. 1. INTRODUCTION Hidden Markov Models (HMMs) are a popular approach for speech recognition. Commonly, a left-to-r...

### Citations

4893 | Neural Networks For Pattern Recognition - Bishop - 1996 |

1106 |
Graphical Models
- Lauritzen
- 1996
(Show Context)
Citation Context ...motivates the addition of data-derived statistical dependencies to an HMM. Section 3 outlines a dependency selection algorithm based on conditional mutual information. Section 4 uses a graphical model=-=[9]-=- to describe the result. Section 5 introduces and derives EM update equations for Gaussian mixture extensions. Section 6 provides WER results on the PHONEBOOK database. 2. LEARNING BOTH STATISTICAL DE... |

744 |
Statistical Methods for Speech Recognition
- Jelinek
- 1997
(Show Context)
Citation Context ...dern speech recognition systems use HMMs to represent model a given probability distribution to an arbitrarily high degree probability distributions. A reasonable approach therefore is to of accuracy =-=[3, 7]-=-. The conditional independence assumptions as- augment an HMM with only those statistical dependencies that sociated with HMMs as they are used in practice, however, have not are found to be missing b... |

496 | Factorial hidden Markov models
- Ghahramani, Jordan
- 1997
(Show Context)
Citation Context ...ve. observation dependencies. One method to increase an HMM’s ability is to increase the number of hidden states[7, 3] or to use a factored hidden-variable 3. BUILDING BMM DEPENDENCIES representation =-=[6]-=-. An alternative method adds explicit fixed dependencies from observations to other observations in the acoustic context, e.g., AR-HMMs [12, 8] and correlation models [13]. And For a given number of h... |

463 | Connectionist Speech Recognition. A Hybrid Approach - Bourlard, Morgan - 1994 |

238 |
Could information theory provide an ecological theory of sensory processing
- Atick
- 1992
(Show Context)
Citation Context ...ratic statistical dependencies. There have been successful attempts to predict neural processing using only environmental Hidden Markov Models (HMMs) are the most common method statistical properties =-=[1]-=-. Indeed, it is these statistical dependencies used in automatic speech recognition systems to model the joint that separate signal from noise and can help distinguish one class probability distributi... |

219 | From HMM’s to segment models: A unified view of stochastic modeling for speech recognition
- Ostendorf, Digalakis, et al.
- 1996
(Show Context)
Citation Context ...observations to other observations in the acoustic context, e.g., AR-HMMs [12, 8] and correlation models [13]. And For a given number of hidden-variable states, the degree to which in segmental models=-=[10]-=-, a hidden state corresponds to a segment a hidden variable does not contain contextual information can be trajectory that defines local statistics. measured using conditional mutual information. The ... |

76 |
Linear predictive hidden Markov models and the speech signal
- Poritz
- 1982
(Show Context)
Citation Context ... hidden-variable 3. BUILDING BMM DEPENDENCIES representation [6]. An alternative method adds explicit fixed dependencies from observations to other observations in the acoustic context, e.g., AR-HMMs =-=[12, 8]-=- and correlation models [13]. And For a given number of hidden-variable states, the degree to which in segmental models[10], a hidden state corresponds to a segment a hidden variable does not contain ... |

76 | A review of large-vocabulary continuous-speech recognition - Young - 1996 |

64 | Mixed Memory Markov Models: Decomposing Complex Stochastic Processes as Mixtures of Simpler Ones - Saul, Jordan - 1999 |

60 |
PhoneBook: A phonetically-rich isolatedword telephone-speech database
- Pitrelli, Fong, et al.
- 1995
(Show Context)
Citation Context ...iable. Previous results [3] showed WER improvements on a small-vocabulary isolated-word speech database using wholeword models. In this paper, preliminary reports show WER reductions for the PHONEBOOK=-=[11]-=- database relative to a baseline HMM. To Appear IEEE ICASSP 1999. 1 X1:(t?1) setfX1;:::;XNgandX<t= the quantity of additional information X<tprovides aboutXtnot already provided byQt, whereQtis the hi... |

44 |
Explicit Time Correlation in Hidden Markov Models for Speech Recognition
- Wellekens
- 1987
(Show Context)
Citation Context ...M DEPENDENCIES representation [6]. An alternative method adds explicit fixed dependencies from observations to other observations in the acoustic context, e.g., AR-HMMs [12, 8] and correlation models =-=[13]-=-. And For a given number of hidden-variable states, the degree to which in segmental models[10], a hidden state corresponds to a segment a hidden variable does not contain contextual information can b... |

42 |
A linear predictive HMM for vector-valued observations with applications to speech recognition
- Kenny, Lennig, et al.
- 1990
(Show Context)
Citation Context ... hidden-variable 3. BUILDING BMM DEPENDENCIES representation [6]. An alternative method adds explicit fixed dependencies from observations to other observations in the acoustic context, e.g., AR-HMMs =-=[12, 8]-=- and correlation models [13]. And For a given number of hidden-variable states, the degree to which in segmental models[10], a hidden state corresponds to a segment a hidden variable does not contain ... |

35 | A model of articulatory dynamics and control - Coker - 1976 |

33 | A multi-channel/multi-speaker articulatory database for continuous speech recognition research - Wrench |

28 | Hybrid HMM/ANN systems for training independent tasks
- Dupont, Bourlard, et al.
- 1997
(Show Context)
Citation Context ...word, telephonespeech database[11]. All data is represented using 12 MFCCs plusc0 and deltas resulting in ad=26 element feature vector sampled every 10ms. The training and test sets are as defined in =-=[5]-=-. Test words do not occur in the training vocabulary, so test word models are constructed using phone-models learned during training. Strictly left-to-right transition matrices were used except for an... |

28 | Combining articulatory and acoustic information for speech recognition in noisy and reverberant environments - Kirchhoff - 1998 |

24 | Hidden-Articulator Markov Models: Performance Improvements and Robustness - Richardson, Bilmes, et al. - 2000 |

21 | Factorial HMMs for acoustic modeling - Logan, Moreno - 1998 |

19 | Data-driven extensions to hmm statistical dependencies
- Bilmes
- 1998
(Show Context)
Citation Context ...dern speech recognition systems use HMMs to represent model a given probability distribution to an arbitrarily high degree probability distributions. A reasonable approach therefore is to of accuracy =-=[3, 7]-=-. The conditional independence assumptions as- augment an HMM with only those statistical dependencies that sociated with HMMs as they are used in practice, however, have not are found to be missing b... |

18 | An HMM-based speech recognizer using overlapping articulatory features - Erler, Freeman - 1996 |

14 | Phonetic Classification and Recognition Using HMM Representation of Overlapping Articulatory Features for all classes of English sounds,” ICASSP - Deng, Sun - 1994 |

10 | Pitch-synchronous frame-by-frame and segment-based articulatory analysis by synthesis - Gupta, Schroeter - 1993 |

9 | Applied Phonetics: The Sounds of American English second edition - Edwards - 1997 |

9 | Factorial HMMs for acoustic modeling,” ICASSP - Logan, Moreno - 1998 |

5 |
A gentle tutorial of the EM algorithm and its application for Gaussian mixture and hidden Markov models
- Bilmes
- 1998
(Show Context)
Citation Context ...e update rule for 1 becomes Xm;v;t?1 2 log(jΣmvj)+(wmvt?m)0Σ?1 :Xm;v;ilogp(mjv;Θ)mvt? mv(wmvt?m)mvt Xv;tlogp(vjzt;Θ)p(vjzt;Θg) Xmp(mjv;Θ)?1! which has the same form as the usual Gaussian mixture case =-=[2, 7]-=-. Equation 2 can be optimized by introducing a Lagrange multiplier Equation 3 can be optimized by noting thatvis assumed independent ofxgivenz: =MXm=1VXv=1TXt=1 wheremvt=p(m;vjxt;zt;Θg), Θgare wherevt... |

2 | PhoneBook: A phonetically-rich iso- telephone speech database” ICASSP - Pitrelli, Fong, et al. - 1995 |

1 | PhoneBook: A phonetically-rich iso1.10 - Pitrelli, Fong, et al. - 1995 |