seba3y commited on
Commit
46dc368
1 Parent(s): 9bdfa19

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +53 -0
README.md CHANGED
@@ -62,6 +62,59 @@ configs:
62
  ---
63
  # speechocean762: A non-native English corpus for pronunciation scoring task
64
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
65
  ## Introduction
66
  Pronunciation scoring is a crucial technology in computer-assisted language learning (CALL) systems. The pronunciation quality scores might be given at phoneme-level, word-level, and sentence-level for a typical pronunciation scoring task.
67
 
 
62
  ---
63
  # speechocean762: A non-native English corpus for pronunciation scoring task
64
 
65
+ ## How to use?
66
+
67
+ you can load data using
68
+
69
+ ```py
70
+ speechocean762_dataset = load_dataset('seba3y/speechocean762')
71
+ ```
72
+ ```py
73
+ >> speechocean762_dataset
74
+ DatasetDict({
75
+ train: Dataset({
76
+ features: ['spk', 'age', 'gender', 'utt_name', 'audio', 'utt_text', 'utt_accuracy', 'utt_completeness', 'utt_fluency', 'utt_prosodic', 'utt_total', 'words', 'words_accuracy', 'words_stress', 'words_total', 'phones', 'phones_godness'],
77
+ num_rows: 2500
78
+ })
79
+ test: Dataset({
80
+ features: ['spk', 'age', 'gender', 'utt_name', 'audio', 'utt_text', 'utt_accuracy', 'utt_completeness', 'utt_fluency', 'utt_prosodic', 'utt_total', 'words', 'words_accuracy', 'words_stress', 'words_total', 'phones', 'phones_godness'],
81
+ num_rows: 2500
82
+ })
83
+ })
84
+ ```
85
+ Features are ordered as following
86
+ 1- Demographic featurs: `'spk', 'age', 'gender', 'utt_name'`
87
+ 2- Sentence-level featurs: `'audio', 'utt_text', 'utt_accuracy', 'utt_completeness', 'utt_fluency', 'utt_prosodic', 'utt_total'`
88
+ 3- Word-level featurs: `'words', 'words_accuracy', 'words_stress', 'words_total'`
89
+ 4- Phoneme-level featurs: `'phones', 'phones_godness'`
90
+
91
+ ```py
92
+ >> speechocean762_dataset['train'][0]
93
+ ```
94
+ ```py
95
+ {'spk': 1,
96
+ 'age': 6,
97
+ 'gender': 'm',
98
+ 'utt_name': 10011,
99
+ 'audio': {'path': '000010011.WAV',
100
+ 'array': array([-9.46044922e-04, -2.38037109e-03, -1.31225586e-03, ...,
101
+ -9.15527344e-05, 3.05175781e-04, -2.44140625e-04]),
102
+ 'sampling_rate': 16000},
103
+ 'utt_text': 'WE CALL IT BEAR',
104
+ 'utt_accuracy': 8,
105
+ 'utt_completeness': 10.0,
106
+ 'utt_fluency': 9,
107
+ 'utt_prosodic': 9,
108
+ 'utt_total': 8,
109
+ 'words': "['WE', 'CALL', 'IT', 'BEAR']",
110
+ 'words_accuracy': '[10, 10, 10, 6]',
111
+ 'words_stress': '[10, 10, 10, 10]',
112
+ 'words_total': '[10, 10, 10, 6]',
113
+ 'phones': "[['W', 'IY0'], ['K', 'AO0', 'L'], ['IH0', 'T'], ['B', 'EH0', 'R']]",
114
+ 'phones_godness': '[[2.0, 2.0], [2.0, 1.8, 1.8], [2.0, 2.0], [2.0, 1.0, 1.0]]'}
115
+ ```
116
+ For word-level features, the 'words' in each sample is a list containing words, while 'words_accuracy', 'words_stress', and 'words_total' are lists of the same length as the words. The mapping is such that the first word corresponds to the first value in 'words_accuracy', and so on. On the other hand, for phoneme-level features, the 'phones' in each sample is a 2D list, with each sublist corresponding to a single word
117
+
118
  ## Introduction
119
  Pronunciation scoring is a crucial technology in computer-assisted language learning (CALL) systems. The pronunciation quality scores might be given at phoneme-level, word-level, and sentence-level for a typical pronunciation scoring task.
120