Added more details about the dataset
#6
by
Aliskin
- opened
README.md
CHANGED
@@ -36,4 +36,72 @@ tags:
|
|
36 |
|
37 |
### Dataset Summary
|
38 |
|
39 |
-
VoxDIY RusNews is
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
36 |
|
37 |
### Dataset Summary
|
38 |
|
39 |
+
VoxDIY RusNews is the first publicly available large-scale dataset of crowdsourced audio transcriptions in Russian language.
|
40 |
+
The dataset was constructed by annotating audio recordings of Russian sentences from news domain on [Toloka crowdsourcing platform](https://toloka.ai).
|
41 |
+
VoxDIY RusNews consists of 3091 instances having around 21K annotations obtained from crowd workers.
|
42 |
+
|
43 |
+
### Supported Tasks and Leaderboards
|
44 |
+
Aggregation of crowd transcriptions.
|
45 |
+
|
46 |
+
### Languages
|
47 |
+
Russian
|
48 |
+
|
49 |
+
## Dataset Structure
|
50 |
+
|
51 |
+
### Data Instances
|
52 |
+
|
53 |
+
A data instance contains a url to the audio recording, a list of transcriptions along with the corresponding performers identifiers and
|
54 |
+
ground truth. For each data instance, seven crowdsourced transcriptions are provided.
|
55 |
+
|
56 |
+
```
|
57 |
+
{'task': 'https://tlk.s3.yandex.net/annotation_tasks/russian/1003.wav',
|
58 |
+
'transcriptions': 'в список так же попали мэрлин монро джон ленон и альберт эйнштейн | в список также попали мерлин монро джон ленон и альберт энштейн | в список также попали мерилин монро джон леннон и альберт энтштейн | в список также попали мэрилин монро джон леннон и альберт эпштейн | в список также попали мэрилин монро джон леннон и альберт эйнштейн | в список так же попали мерелин монро джон ленон и альберт нштейн | в список также попали мэрилин монро джон леннон и альберт эйнштейн',
|
59 |
+
'performers': '1743 | 784 | 1014 | 1572 | 744 | 2187 | 1208',
|
60 |
+
'gt': 'в список также попали мэрилин монро джон леннон и альберт эйнштейн'}
|
61 |
+
```
|
62 |
+
|
63 |
+
### Data Fields
|
64 |
+
|
65 |
+
* task: a string containing a url of the audio recording
|
66 |
+
* transcriptions: a list of the crowdsourced transcriptions separated by '|'
|
67 |
+
* performers: the corresponding performers' identifiers.
|
68 |
+
* gt: ground truth transcription
|
69 |
+
|
70 |
+
## Dataset Creation
|
71 |
+
|
72 |
+
### Source Data
|
73 |
+
|
74 |
+
The audio recordings were obtained using a [speech synthesis tool](https://cloud.yandex.com/en-ru/services/speechkit).
|
75 |
+
The source sentences come from the Russian test set of the machine translation shared task executed as a part of the
|
76 |
+
Eights and Ninth Workshops on Statistical Machine Translation ([WMT 2013](https://www.statmt.org/wmt13/) and [WMT 2014](https://www.statmt.org/wmt14/)).
|
77 |
+
|
78 |
+
### Annotations
|
79 |
+
|
80 |
+
Annotation was done on [Toloka crowdsourcing platform](https://toloka.ai) with overlap of 7 (that is, each task was performed by 7 annotators).
|
81 |
+
|
82 |
+
Only annotators who self-reported the knowledge of Russian had access to the annotation task.
|
83 |
+
Additionally, annotators had to pass *Entrance Exam*. For this, we ask all incoming eligible workers to annotate ten audio
|
84 |
+
recordings. We then compute our target metric — Word Error Rate (WER) — on these recordings and accept to the main task all workers
|
85 |
+
who achieve WER of 40% or less (the smaller the value of the metric, the higher the quality of annotation).
|
86 |
+
|
87 |
+
The Toloka crowdsourcing platform associates workers with unique identifiers and returns these identifiers to the requester.
|
88 |
+
To further protect the data, we additionally encode each identifier with an integer that is eventually reported in our released datasets.
|
89 |
+
|
90 |
+
See more details in the [paper](https://arxiv.org/pdf/2107.01091.pdf).
|
91 |
+
|
92 |
+
### Citation Information
|
93 |
+
|
94 |
+
```
|
95 |
+
@inproceedings{CrowdSpeech,
|
96 |
+
author = {Pavlichenko, Nikita and Stelmakh, Ivan and Ustalov, Dmitry},
|
97 |
+
title = {{CrowdSpeech and Vox~DIY: Benchmark Dataset for Crowdsourced Audio Transcription}},
|
98 |
+
year = {2021},
|
99 |
+
booktitle = {Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks},
|
100 |
+
eprint = {2107.01091},
|
101 |
+
eprinttype = {arxiv},
|
102 |
+
eprintclass = {cs.SD},
|
103 |
+
url = {https://openreview.net/forum?id=3_hgF1NAXU7},
|
104 |
+
language = {english},
|
105 |
+
pubstate = {forthcoming},
|
106 |
+
}
|
107 |
+
```
|