mirix commited on
Commit
20a693e
1 Parent(s): c077029

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +78 -0
README.md CHANGED
@@ -1,3 +1,81 @@
1
  ---
2
  license: mit
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
  license: mit
3
+ task_categories:
4
+ - audio-classification
5
+ language:
6
+ - en
7
+ tags:
8
+ - SER
9
+ - Speech Emotion Recognition
10
+ - Speech Emotion Classification
11
+ - Audio Classification
12
+ - Audio
13
+ - Emotion
14
+ - Emo
15
+ - Speech
16
+ - Mosei
17
+ pretty_name: messiah
18
+ size_categories:
19
+ - 10K<n<100K
20
  ---
21
+
22
+
23
+ DATASET DESCRIPTION
24
+
25
+ The MESSAIH dataset is a fork of [CMU MOSEI](http://multicomp.cs.cmu.edu/resources/cmu-mosei-dataset/).
26
+
27
+ Unlike its parent, MESSAIH is indended for unimodal model development and focusses exclusively on audio classification, more specifically, Speech Emotion Recognition (SER).
28
+
29
+ Of course, it can be used for bimodal classification by transcribing each audio track.
30
+
31
+ MESSAIH currently contains 13,234 speech samples annotated according to the [CMU MOSEI](https://aclanthology.org/P18-1208/) scheme:
32
+
33
+ > Each sentence is annotated for sentiment on a [-3,3] Likert scale of:
34
+ > [−3: highly negative, −2 negative, −1 weakly negative, 0 neutral, +1 weakly positive, +2 positive, +3 highly positive].
35
+ > Ekman emotions of {happiness, sadness, anger, fear, disgust, surprise}
36
+ > are annotated on a [0,3] Likert scale for presence of emotion
37
+ > x: [0: no evidence of x, 1: weakly x, 2: x, 3: highly x].
38
+
39
+ The dataset is provided as a parquet file.
40
+
41
+ To facilitate inspection, a csv file is also provided, but it does not contain the audio arrays.
42
+
43
+ If you train a model on this dataset, you would make us very happy by letting us know.
44
+
45
+
46
+ UNPACKING THE DATASET
47
+
48
+ A sample Python script (check the top of the script for the requirements) is also provided for illustrative purposes.
49
+
50
+ The script reads the parquet file and produces the following:
51
+
52
+ 1. A CSV file with file names and MOSEI values (columns names are self-explanatory). This file is identical to the one already provided.
53
+
54
+ 2. A folder named "wavs" containing the audio samples.
55
+
56
+
57
+ LEGAL CONSIDERATIONS
58
+
59
+ Note that producing the wav files might (or might not) constitute copyright infringement as well as a violation of Google's Terms of Service.
60
+
61
+ Instead, researchers are encouraged to use the numpy arrays contained in the last column of the dataset ("wav2numpy") directly, without actually extracting any playable audio.
62
+
63
+ That, I believe, may keep us in the greyzone.
64
+
65
+
66
+ CAVEATS
67
+
68
+ As one can appreciate from the charts contained in the "chart" folder, the dataset is biased towards "positive" emotions, namely happiness.
69
+
70
+ Certain emotions such as fear may be underrepresented, not only in terms of number of occurences, but, more problematically, in terms of "intensity".
71
+
72
+ MOSEI is considered a natural or spontaneous emotion dataset (as opposed to an actored or scripted one) showcasing "genuine" emotions.
73
+
74
+ However, keep in mind that MOSEI was curated from a popular social network and social networks are notoriously abundant in fake emotions.
75
+
76
+ Moreover, certain emotions may be intrinsically more difficult to detect than others, even from a human perspective.
77
+
78
+ Yet, MOSEI is possibly one of the best datasets in the public domain.
79
+
80
+ Also note that the original [MOSEI](http://immortal.multicomp.cs.cmu.edu/CMU-MOSEI/labels/) contains nearly twice as many entries as MESSAIH does.
81
+