Articulatory feature recognition using dynamic Bayesian networks

Question

1. What are the contributions mentioned in the paper "Articulatory feature recognition using dynamic bayesian networks" ?

2. What are the future works mentioned in the paper "Articulatory feature recognition using dynamic bayesian networks" ?

3. What are the common models used in ASR?

4. What is the definition of a Bayesian network?

Accepted Answer

This paper describes the use of dynamic Bayesian networks for the task of articulatory feature recognition.. The authors show that by modeling the dependencies between a set of 6 multi-leveled articulatory features, recognition accuracy is increased over an equivalent system in which features are considered independent.

Accepted Answer

Future work will use embedded training in which the sequence of features is specified but not the timings of transitions.. This approach will allow asynchronous feature changes, though in the absence of suitably detailed articulatory feature labels it is not clear how to evaluate such a system directly.

Accepted Answer

The models studied include artificial neural networks (ANN) [5, 6, 7], hidden Markov models (HMM)[5], linear dynamic models (LDM) [4], and dynamic Bayesian networks (DBN) [3, 8].

Accepted Answer

A Bayesian network exploits missing edges (implying conditional independence) to factor the joint distribution of all random variables into a set of simpler probability distributions.

Accepted Answer

With a node representing each of the 6 feature groups, and given that DBNs are directed acyclic graphs, there can be at most 5+4+3+2+1 = 15 within-frame edges.

Accepted Answer

The acoustic waveform was parameterized as 12 Mel-frequency cepstral coefficients (MFCC) and energy, calculated every 10ms within 25ms windows.

Accepted Answer

During training, both acoustic and articulatory features are observed, and recognition uses a Viterbi search to find the most likely sequence of feature values given acoustic input.

Accepted Answer

For the independent models, there were 3032 combinations found in the recognition output, reduced to 1084 where features were forced to synchronize use of the silence model.

Accepted Answer

The articulatory features can be derived from measured articulation [2, 3, 4], or generated from existing labels according to linguistic knowledge [5, 6, 7].

Accepted Answer

One approach is the class of hidden feature models described by [8], in which articulatory features were used for feature-based factorization of the observation space.

Accepted Answer

Forcing common silence/non-silence decisions increases the average feature accuracy from 80.1% to 81.1%, and the percentage of frames where all features are correct simultaneously is increased from 45.2% to 50.8%.

Accepted Answer

The conditional entropy H(Y |X) is given byH(Y |X) = − ∑ x ∑ y P (x, y) log2 P (y|x) (4)and indicates the uncertainty in Y given knowledge of X .

Accepted Answer

The mutual information I(X;Y ) between a pair of discrete random variables X and Y is calculated as:I(X;Y ) = ∑ x ∑ y P (x, y) log2 P (x, y) P (x)P (y) (3)and gives a measure of how much information one variable provides about the other.

Accepted Answer

The authors intend to address this and develop an implementation using distributions specific to feature combinations where possible, backing off to product models where training data is limited.

Accepted Answer

The approach taken in this work is to commence with a baseline model in which the 6 feature groups are modeled as independent, add edges one at a time and evaluate each model on validation data.

Articulatory feature recognition using dynamic Bayesian networks

Chat with Paper

AI Agents for this Paper

Most frequently asked questions

1. What are the contributions mentioned in the paper "Articulatory feature recognition using dynamic bayesian networks" ?

2. What are the future works mentioned in the paper "Articulatory feature recognition using dynamic bayesian networks" ?

3. What are the common models used in ASR?

4. What is the definition of a Bayesian network?

5. How many within-frame edges can be added?

6. How many MFCCs were used in the acoustic waveform?

7. What is the way to find the likely sequence of features?

8. How many combinations were found in the recognition output?

9. What is the common way to model the articulatory features of a speech?

10. What is the way to factor the observation space?

11. How many frames are correct when a feature is forced to take silence?

12. What is the conditional entropy of a feature?

13. What is the mutual information between a pair of discrete random variables X and Y?

14. What is the plan to address this problem?

15. How many features are modeled as independent?

Figures

Citations

Statistical Parametric Speech Synthesis

Speech production knowledge in automatic speech recognition.

Modeling coarticulation in EMG-based continuous speech recognition

Articulatory feature recognition using dynamic Bayesian networks

Articulatory Information for Noise Robust Speech Recognition

References

The Sound Pattern of English

Bayesian networks and decision graphs

A course in phonetics

Articulatory phonology: an overview.

Robust speech recognition using articulatory information

Related Papers (5)

Asynchronous Articulatory Feature Recognition Using Dynamic Bayesian Networks

Robust speech recognition using articulatory information

Moving beyond the 'beads-on-a-string' model of speech

Articulatory and Spectrum Information Fusion Based on Deep Recurrent Neural Networks

A factored conditional random field model for articulatory feature forced transcription