File size: 13,786 Bytes
5114290 c5b1ba3 2405604 c5b1ba3 aab02c6 2405604 aab02c6 1393f38 5114290 d74712c a231e7a d74712c d0a8d39 d74712c a231e7a 5114290 d74712c d0a8d39 d74712c d0a8d39 d74712c d0a8d39 d74712c d0a8d39 d74712c d0a8d39 d74712c d0a8d39 7efb28c 5114290 7efb28c 990c5e1 7efb28c 990c5e1 7efb28c 5114290 990c5e1 c5b1ba3 9bdfa19 46dc368 c31c3b6 46dc368 3c2ce35 46dc368 3c2ce35 46dc368 3c2ce35 46dc368 e4699de 46dc368 e4699de 46dc368 9bdfa19 c31c3b6 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 |
---
language:
- en
license: apache-2.0
size_categories:
- 1K<n<10K
task_categories:
- audio-classification
- automatic-speech-recognition
pretty_name: ' '
tags:
- pronunciation-scoring
- L1-Mandarin
- L2-English
dataset_info:
features:
- name: spk
dtype: string
- name: age
dtype: string
- name: gender
dtype: string
- name: utt_name
dtype: string
- name: audio
dtype:
audio:
sampling_rate: 16000
- name: utt_text
dtype: string
- name: utt_accuracy
dtype: int64
- name: utt_completeness
dtype: float64
- name: utt_fluency
dtype: int64
- name: utt_prosodic
dtype: int64
- name: utt_total
dtype: int64
- name: words
sequence: string
- name: words_accuracy
sequence: int64
- name: words_stress
sequence: int64
- name: words_total
sequence: int64
- name: phones
sequence:
sequence: string
- name: phones_godness
sequence:
sequence: float64
- name: duration
dtype: float64
splits:
- name: train
num_bytes: 333075617.5
num_examples: 2500
- name: test
num_bytes: 311790040.5
num_examples: 2500
download_size: 611757634
dataset_size: 644865658.0
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: test
path: data/test-*
---
# speechocean762: A non-native English corpus for pronunciation scoring task
## How to use?
you can load data using
```py
speechocean762_dataset = load_dataset('seba3y/speechocean762')
```
```py
>> speechocean762_dataset
DatasetDict({
train: Dataset({
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'],
num_rows: 2500
})
test: Dataset({
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'],
num_rows: 2500
})
})
```
Features are ordered as following:
1- Demographic featurs: `'spk', 'age', 'gender', 'utt_name'`
2- Sentence-level featurs: `'audio', 'utt_text', 'utt_accuracy', 'utt_completeness', 'utt_fluency', 'utt_prosodic', 'utt_total'`
3- Word-level featurs: `'words', 'words_accuracy', 'words_stress', 'words_total'`
4- Phoneme-level featurs: `'phones', 'phones_godness'`
```py
>> speechocean762_dataset['train'][0]
```
```py
{'spk': '0001',
'age': 6,
'gender': 'm',
'utt_name': '000010011',
'audio': {'path': '/content/speechocean762/WAVE/SPEAKER0001/000010011.WAV',
'array': array([-9.46044922e-04, -2.38037109e-03, -1.31225586e-03, ...,
-9.15527344e-05, 3.05175781e-04, -2.44140625e-04]),
'sampling_rate': 16000},
'utt_text': 'WE CALL IT BEAR',
'utt_accuracy': 8,
'utt_completeness': 10.0,
'utt_fluency': 9,
'utt_prosodic': 9,
'utt_total': 8,
'words': "['WE', 'CALL', 'IT', 'BEAR']",
'words_accuracy': '[10, 10, 10, 6]',
'words_stress': '[10, 10, 10, 10]',
'words_total': '[10, 10, 10, 6]',
'phones': "[['W', 'IY0'], ['K', 'AO0', 'L'], ['IH0', 'T'], ['B', 'EH0', 'R']]",
'phones_godness': '[[2.0, 2.0], [2.0, 1.8, 1.8], [2.0, 2.0], [2.0, 1.0, 1.0]]'}
```
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
## Introduction
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.
This corpus aims to provide a free public dataset for the pronunciation scoring task.
Key features:
* It is available for free download for both commercial and non-commercial purposes.
* The speaker variety encompasses young children and adults.
* The manual annotations are in multiple aspects at sentence-level, word-level and phoneme-level.
This corpus consists of 5000 English sentences. All the speakers are non-native, and their mother tongue is Mandarin. Half of the speakers are Children, and the others are adults. The information of age and gender are provided.
Five experts made the scores. To avoid subjective bias, each expert scores independently under the same metric.
## The scoring metric
The experts score at three levels: phoneme-level, word-level, and sentence-level.
### Phoneme level
Score the pronunciation goodness of each phoneme within the words.
Score range: 0-2
* 2: pronunciation is correct
* 1: pronunciation is right but has a heavy accent
* 0: pronunciation is incorrect or missed
### Word level
Score the accuracy and stress of each word's pronunciation.
#### Accuracy
Score range: 0 - 10
* 10: The pronunciation of the word is perfect
* 7-9: Most phones in this word are pronounced correctly but have accents
* 4-6: Less than 30% of phones in this word are wrongly pronounced
* 2-3: More than 30% of phones in this word are wrongly pronounced. In another case, the word is mispronounced as some other word. For example, the student mispronounced the word "bag" as "bike"
* 1: The pronunciation is hard to distinguish
* 0: no voice
#### Stress
Score range: {5, 10}
* 10: The stress is correct, or this is a mono-syllable word
* 5: The stress is wrong
### Sentence level
Score the accuracy, fluency, completeness and prosodic at the sentence level.
#### Accuracy
Score range: 0 - 10
* 9-10: The overall pronunciation of the sentence is excellent, with accurate phonology and no obvious pronunciation mistakes
* 7-8: The overall pronunciation of the sentence is good, with a few pronunciation mistakes
* 5-6: The overall pronunciation of the sentence is understandable, with many pronunciation mistakes and accent, but it does not affect the understanding of basic meanings
* 3-4: Poor, clumsy and rigid pronunciation of the sentence as a whole, with serious pronunciation mistakes
* 0-2: Extremely poor pronunciation and only one or two words are recognizable
#### Completeness
Score range: 0.0 - 1.0
The percentage of the words with good pronunciation.
#### Fluency
Score range: 0 - 10
* 8-10: Fluent without noticeable pauses or stammering
* 6-7: Fluent in general, with a few pauses, repetition, and stammering
* 4-5: the speech is a little influent, with many pauses, repetition, and stammering
* 0-3: intermittent, very influent speech, with lots of pauses, repetition, and stammering
#### Prosodic
Score range: 0 - 10
* 9-10: Correct intonation at a stable speaking speed, speak with cadence, and can speak like a native
* 7-8: Nearly correct intonation at a stable speaking speed, nearly smooth and coherent, but with little stammering and few pauses
* 5-6: Unstable speech speed, many stammering and pauses with a poor sense of rhythm
* 3-4: Unstable speech speed, speak too fast or too slow, without the sense of rhythm
* 0-2: Poor intonation and lots of stammering and pauses, unable to read a complete sentence
## Data structure
The following tree shows the file structure of this corpus on [github](https://github.com/jimbozhang/speechocean762):
```
├── scores.json
├── scores-detail.json
├── train
│ ├── spk2age
│ ├── spk2gender
│ ├── spk2utt
│ ├── text
│ ├── utt2spk
│ └── wav.scp
├── test
│ ├── spk2age
│ ├── spk2gender
│ ├── spk2utt
│ ├── text
│ ├── utt2spk
│ └── wav.scp
└── WAVE
├── SPEAKER0001
│ ├── 000010011.WAV
│ ├── 000010035.WAV
│ ├── ...
│ └── 000010173.WAV
├── SPEAKER0003
│ ├── 000030012.WAV
│ ├── 000030024.WAV
│ ├── ...
│ └── 000030175.WAV
└── SPEAKER0005
├── 000050003.WAV
├── 000050010.WAV
├── ...
└── 000050175.WAV
```
There are two datasets: `train` and `test`, and both are in Kaldi's data directory style.
The scores are stored in `scores.json`. Here is an example:
```
{
"000010011": { # utt-id
"text": "WE CALL IT BEAR", # transcript text
"accuracy": 8, # sentence-level accuracy score
"completeness": 10.0, # sentence-level completeness score
"fluency": 9, # sentence-level fluency score
"prosodic": 9, # sentence-level prosodic score
"total": 8, # sentence-level total score
"words": [
{
"accuracy": 10, # word-level accuracy score
"stress": 10, # word-level stress score
"total": 10, # word-level total score
"text": "WE", # the word text
"phones": "W IY0", # phones of the word
"phones-accuracy": [2.0, 2.0] # phoneme-level accuracy score
},
{
"accuracy": 10,
"stress": 10,
"total": 10,
"text": "CALL",
"phones": "K AO0 L",
"phones-accuracy": [2.0, 1.8, 1.8]
},
{
"accuracy": 10,
"stress": 10,
"total": 10,
"text": "IT",
"phones": "IH0 T",
"phones-accuracy": [2.0, 2.0]
},
{
"accuracy": 6,
"stress": 10,
"total": 6,
"text": "BEAR",
"phones": "B EH0 R",
"phones-accuracy": [2.0, 1.0, 1.0]
}
]
},
...
}
```
For the phones with an accuracy score lower than 0.5, an extra "mispronunciations" block indicates which phoneme the current phone was actually pronounced.
An example:
```
{
"text": "LISA",
"accuracy": 5,
"phones": ["L", "IY1", "S", "AH0"],
"phones-accuracy": [0.4, 2, 2, 1.2],
"mispronunciations": [
{
"canonical-phone": "L",
"index": 0,
"pronounced-phone": "D"
}
],
"stress": 10,
"total": 6
}
```
The file `scores.json` is processed from `scores-detail.json`.
The two JSON files are almost the same, but `scores-detail.json` has the five experts' original scores, while the scores of scores.json were the average or median scores.
An example item in `scores-detail.json`:
```
{
"000010011": {
"text": "WE CALL IT BEAR",
"accuracy": [7.0, 9.0, 8.0, 8.0, 9.0],
"completeness": [1.0, 1.0, 1.0, 1.0, 1.0],
"fluency": [10.0, 9.0, 8.0, 8.0, 10.0],
"prosodic": [10.0, 9.0, 7.0, 8.0, 9.0],
"total": [7.6, 9.0, 7.9, 8.0, 9.1],
"words": [
{
"accuracy": [10.0, 10.0, 10.0, 10.0, 10.0],
"stress": [10.0, 10.0, 10.0, 10.0, 10.0],
"total": [10.0, 10.0, 10.0, 10.0, 10.0],
"text": "WE",
"ref-phones": "W IY0",
"phones": ["W IY0", "W IY0", "W IY0", "W IY0", "W IY0"]
},
{
"accuracy": [10.0, 8.0, 10.0, 10.0, 8.0],
"stress": [10.0, 10.0, 10.0, 10.0, 10.0],
"total": [10.0, 8.4, 10.0, 10.0, 8.4],
"text": "CALL",
"ref-phones": "K AO0 L",
"phones": ["K AO0 L", "K {AO0} L", "K AO0 L", "K AO0 L", "K AO0 {L}"],
},
{
"accuracy": [10.0, 10.0, 10.0, 10.0, 10.0],
"stress": [10.0, 10.0, 10.0, 10.0, 10.0],
"total": [10.0, 10.0, 10.0, 10.0, 10.0],
"text": "IT",
"ref-phones": "IH0 T",
"phones": ["IH0 T", "IH0 T", "IH0 T", "IH0 T", "IH0 T"]
},
{
"accuracy": [3.0, 7.0, 10.0, 2.0, 6.0],
"stress": [10.0, 10.0, 10.0, 10.0, 10.0],
"phones": ["B (EH0) (R)", "B {EH0} {R}", "B EH0 R", "B (EH0) (R)", "B EH0 [L] R"],
"total": [4.4, 7.6, 10.0, 3.6, 6.8],
"text": "BEAR",
"ref-phones": "B EH0 R"
}
],
},
...
}
```
In `scores-detail.json`, the phoneme-level scores are notated in the following convenient notation:
* for score 2, do not use any symbol
* for score 1, use "{}" symbol
* for score 0, use "()" symbol
* for the inserted phone, use the "[]" symbol
For example, "B (EH) R" means the score of EH is 0 while the scores of B and R are both 2,
"B EH [L] R" mean there is an unexpected phone "L" and the other phones are scored 2.
## Citation
Please cite our paper if you find this work useful:
```bibtex
@inproceedings{zhang2021speechocean762,
title={speechocean762: An Open-Source Non-native English Speech Corpus For Pronunciation Assessment},
author={Zhang, Junbo and Zhang, Zhiwen and Wang, Yongqing and Yan, Zhiyong and Song, Qiong and Huang, Yukai and Li, Ke and Povey, Daniel and Wang, Yujun},
booktitle={Proc. Interspeech 2021},
year={2021}
}
``` |