Learning mixture models for speech recognition

1
Learning mixture models for speech recognition

• Rohit Prabhavalkar,Prateeti Mohapatra

2
Outline

• Introduction
• Learning mixture models using EM with k-means
• The experiment
• Results
• Future Work

3
Speech recognition

• Given acoustic inputs (converted into feature
  vectors eg. PLP or MFCCs)?
• Find argmax P(QX)?
• Q
• Where,
• Q is most likely phone sequence
• X is the input sequence of feature vectors

4
Conventional techniques for acoustic modeling

• Generative models, such as HMMs, estimate p(xq)?
• Usually modeled by a mixture of gaussians for
  each phone
• Conventionally this is done by EM, with k-means
  initialization

5
Problems with conventional techniques

• EM converges to local maxima
• 'Good' initialization for k-means is critical
• It is usually hard to estimate the number of
  gaussians to use

Image taken from http//home.dei.polimi.it/matte
ucc/Clustering/tutorial_html/AppletKM.html
6
Problems with conventional techniques

• EM converges to local maxima
• 'Good' initializations for k-means is critical
• It is usually hard to estimate the number of
  gaussians to use

Image taken from http//home.dei.polimi.it/matte
ucc/Clustering/tutorial_html/AppletKM.html
7
The algorithm

• Uses eigenvectors of the Kernel Matrix to
  estimate parameters of a Gaussian Mixture Model

• Estimates number of mixtures to use, along
• with means and covariances of each mixture
• Choosing the bandwidth 'w' is a challenge

8
CRFs

• Are discriminative models
• Do not make independence assumptions about the
  input unlike HMMs
• Have been shown to achieve better performance
  than HMMs for speech recogntion

9
The experiments

• Performed phone recognition on the TIMIT dataset
• Used a Tandem System as the baseline
• Test system was based on a CRF
• Needed to randomly sample from the training set
  due to time constraints

10
Results
11
Future work

• Lack of time prevented us from experimenting with
  different bandwidth values
• Could increase number of samples used in the
  algorithm
• Could use the values predicted by algorithm as
  initialization for EM