|
Introduction
Data
& Labeling
Feature
Extraction
Models
Application
Results
|
Summary of Results:
The following values indicate percentage of correctly classified
messages given all the available data. We have acquired these percentages using leave-one-out cross validation
testing on all the data.
|
model |
Happy(40) |
Sad(27) |
Calm(84) |
Excited(77) |
Urgent(46) |
Not Urgent(140) |
Formal(98) |
Informal(150) |
| GMM
N=1 Diagonal |
75 |
44 |
76 |
52 |
35 |
70 |
73 |
44 |
| GMM
N=1 Full |
87 |
11 |
79 |
68 |
17 |
92 |
60 |
59 |
| GMM
N=2 Diagonal |
65 |
37 |
75 |
67.5 |
37 |
63.5 |
49 |
68 |
| GMM
N=4 Diagonal |
70 |
37 |
74 |
62 |
37 |
68.5 |
64 |
57 |
| GMM
N=2 Full |
NA |
NA |
74 |
57 |
NA |
NA |
37 |
65 |
| HMM N=3 Fully
Connected |
79 |
73 |
67 |
67 |
66 |
76 |
68 |
64 |
* We have experimented with HMMs with different states and saw
that 3 states provided optimal results. Due to the lengthy process
of cross-validation, we did not perform tests on all the data for
these cases.
Conclusions:
- Happy, Sad and Urgent, Not Urgent were not separable using the
global features and GMMs. HMMs were able to discriminate the
data in these cases.
- Both the GMM and HMM were able separate formal and informal,
with HMMs performing slightly better. (5%)
- GMM with 2 mixtures or single mixture with full covariance was able to separate the activation axis
(calm, excited) slightly better than the HMM. This hints at the
fact that formality and activation are somewhat related to
global characteristics of the speech signal , while urgency and
valence are reflected in local transitions.
|
