Datasets:

Languages:
English
ArXiv:
License:
japerez commited on
Commit
b7e8587
·
verified ·
1 Parent(s): 32a2e24

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +63 -0
README.md CHANGED
@@ -8,3 +8,66 @@ language:
8
 
9
  A dataset containing English speech with grammatical errors, along with the corresponding transcriptions. Utterances are synthesized using a
10
  text-to-speech model, whereas the grammatically incorrect texts come from the [C4_200M](https://aclanthology.org/2021.bea-1.4) synthetic dataset.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8
 
9
  A dataset containing English speech with grammatical errors, along with the corresponding transcriptions. Utterances are synthesized using a
10
  text-to-speech model, whereas the grammatically incorrect texts come from the [C4_200M](https://aclanthology.org/2021.bea-1.4) synthetic dataset.
11
+
12
+ ## Introduction
13
+
14
+ The Synthesized English Speech with Grammatical Errors (SESGE) dataset was developed to support the [DeMINT](https://github.com/transducens/demint) project.
15
+ The objective of DeMINT is to develop an intelligent tutoring system that helps non-native English speakers improve their language skills by analyzing and providing
16
+ feedback on the transcripts of their online meetings. As part of this, a system able to transcribe spoken English keeping the original
17
+ grammatical errors intact was essential.
18
+ Existing speech-to-text (STT) models like Whisper tend to correct grammatical errors due to their strong internal language models, making them unsuitable for this task.
19
+ Therefore, SESGE was created to train a custom STT model that could accurately transcribe spoken English with grammatical errors preserved.
20
+
21
+ ## Dataset Creation
22
+
23
+ Given the absence of a suitable dataset for training an error-preserving STT system, DeMINT fine-tuned a Whisper model with data from two primary sources:
24
+
25
+ - [COREFL](https://www.peterlang.com/document/1049094)
26
+ The COREFL dataset consists of essays written by non-native English students with various levels of proficiency.
27
+ While some of these essays have associated audio recordings, the majority do not.
28
+ To expand the audio dataset, we used the [StyleTTS2](https://arxiv.org/abs/2306.07691) text-to-speech model to generate synthetic audio for the remaining texts.
29
+ Multiple voices were used for synthesis to increase the diversity of the dataset.
30
+ COREFL also includes audio directly recorded by students, which introduces natural speech variability and common errors found among L1-Spanish speakers,
31
+ a key demographic for the DeMINT project.
32
+
33
+ - [C4_200M](https://github.com/google-research-datasets/C4_200M-synthetic-dataset-for-grammatical-error-correction)
34
+ The C4_200M dataset contains synthetically generated English sentences with grammatical errors, produced using a corruption model.
35
+ Like with COREFL, StyleTTS2 was employed to synthesize audio from these texts, diversifying the voices to enhance the training set.
36
+ This dataset primarily provides varied sentence structures and error types, although with a limited number of distinct voices.
37
+
38
+ Due to licensing restrictions associated with the COREFL dataset, only the portion derived from the C4_200M dataset is publicly available as part of the
39
+ SESGE dataset. This means that while COREFL data was used during our training, only the C4_200M-based data is included in this dataset.
40
+
41
+ Training samples comprise **28,592** utterances from C4_200M.
42
+
43
+ ## Models
44
+
45
+ Two models were trained on the SESGE dataset by fine-tuning Whisper, enabling error-preserving STT. These models are available on the Hugging Face Hub:
46
+
47
+ - [Error-Preserving Whisper Model](https://huggingface.co/Transducens/error-preserving-whisper)
48
+ - [Error-Preserving Whisper Distilled Model](https://huggingface.co/Transducens/error-preserving-whisper-distilled)
49
+
50
+ Both models have been optimized to transcribe spoken English while retaining grammatical errors, making them suitable for language-learning applications
51
+ where fidelity to spoken errors is essential.
52
+
53
+ ## How to Cite
54
+
55
+ If you use the SESGE dataset, please cite the following paper:
56
+
57
+ ```bibtex
58
+ @inproceedings{demint2024,
59
+ author = {Pérez-Ortiz, Juan Antonio and
60
+ Esplà-Gomis, Miquel and
61
+ Sánchez-Cartagena, Víctor M. and
62
+ Sánchez-Martínez, Felipe and
63
+ Chernysh, Roman and
64
+ Mora-Rodríguez, Gabriel and
65
+ Berezhnoy, Lev},
66
+ title = {{DeMINT}: Automated Language Debriefing for English Learners via {AI}
67
+ Chatbot Analysis of Meeting Transcripts},
68
+ booktitle = {Proceedings of the 13th Workshop on NLP for Computer Assisted Language Learning},
69
+ month = october,
70
+ year = {2024},
71
+ url = {https://aclanthology.org/volumes/2024.nlp4call-1/},
72
+ }
73
+ ```