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. However, there is no public dataset for this task. Researchers use their private datasets in their papers.
This corpus aims to provide a free public dataset for the pronunciation scoring task.
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 experts score at three levels: phoneme-level, word-level, and sentence-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
Score the accuracy and stress of each word's pronunciation.
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
Score range: {5, 10}
- 10: The stress is correct, or this is a mono-syllable word
- 5: The stress is wrong
Score the accuracy, fluency, completeness and prosodic at the sentence level.
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
Score range: 0.0 - 1.0 The percentage of the words with good pronunciation.
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
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
The following tree shows the file structure of this corpus:
├── 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.
Please cite our paper if you find this work useful:
@inproceedings{speechocean762,
title={speechocean762: An Open-Source Non-native English Speech Corpus For Pronunciation Assessment},
booktitle={Proc. Interspeech 2021},
year=2021,
author={Junbo Zhang, Zhiwen Zhang, Yongqing Wang, Zhiyong Yan, Qiong Song, Yukai Huang, Ke Li, Daniel Povey, Yujun Wang}
}