Hugging Face's logo

Datasets:
MushanW
/
ESLTTS

Tasks:
Text-to-Audio
Automatic Speech Recognition
Audio-to-Audio
Modalities:
Audio
Text
Formats:
parquet
Languages:
English
Size:
10K - 100K
ArXiv:
Libraries:
Datasets
Dask
License:
Dataset card Viewer Files Community
ESLTTS
File size: 3,352 Bytes 
951f8cd
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
 
---
language:
- en
license: cc0-1.0
task_categories:
- text-to-audio
- automatic-speech-recognition
- audio-to-audio
- audio-classification
dataset_info:
  features:
  - name: audio
    dtype: audio
  - name: speaker_id
    dtype: string
  - name: transcript
    dtype: string
  - name: native_language
    dtype: string
  - name: subset
    dtype: string
  splits:
  - name: all
    num_bytes: 3179636720.056
    num_examples: 41806
  download_size: 3667597693
  dataset_size: 3179636720.056
configs:
- config_name: default
  data_files:
  - split: all
    path: data/all-*
---

# ESLTTS

The full paper can be accessed here: [arXiv](https://arxiv.org/abs/2404.18094), [IEEE Xplore](https://ieeexplore.ieee.org/document/10508477).

## Dataset Access
You can access this dataset through [Huggingface](https://huggingface.co/datasets/MushanW/ESLTTS) or [Google Driver](https://drive.google.com/file/d/1ChQ_z-TxvKWNUbUMWnbyjM2VY3v2SKEi/view?usp=sharing) or [IEEE Dataport](http://ieee-dataport.org/documents/english-second-language-tts-esltts-dataset).

## Abstract

With the progress made in speaker-adaptive TTS approaches, advanced approaches have shown a remarkable capacity to reproduce the speaker’s voice in the commonly used TTS datasets. However, mimicking voices characterized by substantial accents, such as non-native English speakers, is still challenging.  Regrettably, the absence of a dedicated TTS dataset for speakers with substantial accents inhibits the research and evaluation of speaker-adaptive TTS models under such conditions. To address this gap, we developed a corpus of non-native speakers' English utterances.

We named this corpus “English as a Second Language TTS dataset ” (ESLTTS). The ESLTTS dataset consists of roughly 37 hours of 42,000 utterances from 134 non-native English speakers. These speakers represent a diversity of linguistic backgrounds spanning 31 native languages. For each speaker, the dataset includes an adaptation set lasting about 5 minutes for speaker adaptation, a test set comprising 10 utterances for speaker-adaptive TTS evaluation, and a development set for further research.

## Dataset Structure

```
ESLTTS Dataset/
├─ Malayalam_3/     ------------ {Speaker Native Language}_{Speaker id}
│  ├─ ada_1.flac    ------------ {Subset Name}_{Utterance id}
│  ├─ ada_1.txt     ------------ Transcription for "ada_1.flac"
│  ├─ test_1.flac   ------------ {Subset Name}_{Utterance id}
│  ├─ test_1.txt    ------------ Transcription for "test_1.flac"
│  ├─ dev_1.flac    ------------ {Subset Name}_{Utterance id}
│  ├─ dev_1.txt     ------------ Transcription for "dev_1.flac"
│  ├─ ...
├─ Arabic_3/        ------------ {Speaker Native Language}_{Speaker id}
│  ├─ ada_1.flac    ------------ {Subset Name}_{Utterance id}
│  ├─ ...
├─ ...
```

## Citation
```
@article{wang2024usat,
  author       = {Wenbin Wang and
                  Yang Song and
                  Sanjay K. Jha},
  title        = {{USAT:} {A} Universal Speaker-Adaptive Text-to-Speech Approach},
  journal      = {{IEEE} {ACM} Trans. Audio Speech Lang. Process.},
  volume       = {32},
  pages        = {2590--2604},
  year         = {2024},
  url          = {https://doi.org/10.1109/TASLP.2024.3393714},
  doi          = {10.1109/TASLP.2024.3393714},
}
```