Datasets:

Languages:
English
ArXiv:
License:
galv commited on
Commit
30bd8ef
1 Parent(s): 13ed505

Fill out datasheet some more.

Browse files

Please feel free to consider this preliminary and suggest changes.

Files changed (1) hide show
  1. README.md +42 -9
README.md CHANGED
@@ -113,39 +113,46 @@ See our [paper](https://arxiv.org/abs/2111.09344).
113
 
114
  #### Initial Data Collection and Normalization
115
 
116
- See our [paper](https://arxiv.org/abs/2111.09344).
117
 
118
  #### Who are the source language producers?
119
 
120
- See our [paper](https://arxiv.org/abs/2111.09344).
121
 
122
  ### Annotations
123
 
124
  #### Annotation process
125
 
126
- See our [paper](https://arxiv.org/abs/2111.09344).
127
 
128
  #### Who are the annotators?
129
 
130
- See our [paper](https://arxiv.org/abs/2111.09344).
131
 
132
  ### Personal and Sensitive Information
133
 
134
- See our [paper](https://arxiv.org/abs/2111.09344).
135
 
136
  ## Considerations for Using the Data
137
 
138
  ### Social Impact of Dataset
139
 
140
- See our [paper](https://arxiv.org/abs/2111.09344).
 
 
 
 
 
141
 
142
  ### Discussion of Biases
143
 
144
- See our [paper](https://arxiv.org/abs/2111.09344).
 
 
145
 
146
  ### Other Known Limitations
147
 
148
- [Needs More Information]
149
 
150
  ## Additional Information
151
 
@@ -159,4 +166,30 @@ We provide CC-BY and CC-BY-SA subsets of the dataset.
159
 
160
  ### Citation Information
161
 
162
- [Needs More Information]
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
113
 
114
  #### Initial Data Collection and Normalization
115
 
116
+ Data was downloaded via the archive.org API. No data inference was done.
117
 
118
  #### Who are the source language producers?
119
 
120
+ [Needs More Information]
121
 
122
  ### Annotations
123
 
124
  #### Annotation process
125
 
126
+ No manual annotation is done. We download only source audio with already existing transcripts.
127
 
128
  #### Who are the annotators?
129
 
130
+ For the test and dev sets, we paid native American English speakers to do transcriptions. We do not know the identities of the transcriptionists for data in the training set. For the training set, we have noticed that some transcriptions are likely to be the output of automatic speech recognition systems.
131
 
132
  ### Personal and Sensitive Information
133
 
134
+ Several of our sources are legal and government proceedings, spoken histories, speeches, and so on. Given that these were intended as public documents and licensed as such, it is natural that the involved individuals are aware of this.
135
 
136
  ## Considerations for Using the Data
137
 
138
  ### Social Impact of Dataset
139
 
140
+ The dataset could be used for speech synthesis. However, this requires careful cleaning of the dataset, as background noise is not tolerable for speech synthesis.
141
+
142
+ The dataset could be used for keyword spotting tasks as well. In particular, this is good use case for the non-English audio in the dataset.
143
+
144
+ Our sincere hope is that the large breadth of sources our dataset incorporates reduces existing quality of service issues today, like speech recognition system’s poor understanding of non-native English accents. We cannot think of any unfair treatment that come from using this dataset at this time.
145
+
146
 
147
  ### Discussion of Biases
148
 
149
+ Our data is downloaded from archive.org. As such, the data is biased towards whatever users decide to upload there.
150
+
151
+ Almost all of our data is American accented English.
152
 
153
  ### Other Known Limitations
154
 
155
+ As of version 1.0, a portion of data in the training, test, and dev sets is poorly aligned. Specifically, some words appear in the transcript, but not the audio, or some words appear in the audio, but not the transcript. We are working on it.
156
 
157
  ## Additional Information
158
 
 
166
 
167
  ### Citation Information
168
 
169
+ Please cite:
170
+
171
+ ```
172
+ @article{DBLP:journals/corr/abs-2111-09344,
173
+ author = {Daniel Galvez and
174
+ Greg Diamos and
175
+ Juan Ciro and
176
+ Juan Felipe Cer{\'{o}}n and
177
+ Keith Achorn and
178
+ Anjali Gopi and
179
+ David Kanter and
180
+ Maximilian Lam and
181
+ Mark Mazumder and
182
+ Vijay Janapa Reddi},
183
+ title = {The People's Speech: {A} Large-Scale Diverse English Speech Recognition
184
+ Dataset for Commercial Usage},
185
+ journal = {CoRR},
186
+ volume = {abs/2111.09344},
187
+ year = {2021},
188
+ url = {https://arxiv.org/abs/2111.09344},
189
+ eprinttype = {arXiv},
190
+ eprint = {2111.09344},
191
+ timestamp = {Mon, 22 Nov 2021 16:44:07 +0100},
192
+ biburl = {https://dblp.org/rec/journals/corr/abs-2111-09344.bib},
193
+ bibsource = {dblp computer science bibliography, https://dblp.org}
194
+ }
195
+ ```