polinaeterna HF staff commited on
Commit
900248f
1 Parent(s): 0d6394e

add loading scrip and readme

Browse files
Files changed (3) hide show
  1. README.md +297 -3
  2. dataset_infos.json +0 -0
  3. ml_spoken_words.py +244 -0
README.md CHANGED
@@ -1,3 +1,297 @@
1
- ---
2
- license: cc-by-4.0
3
- ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ annotations_creators:
3
+ - machine-generated
4
+ language_creators:
5
+ - other
6
+ languages:
7
+ - ar
8
+ - as
9
+ - br
10
+ - ca
11
+ - cnh
12
+ - cs
13
+ - cv
14
+ - cy
15
+ - de
16
+ - dv
17
+ - el
18
+ - en
19
+ - eo
20
+ - es
21
+ - et
22
+ - eu
23
+ - fa
24
+ - fr
25
+ - fy-NL
26
+ - ga-IE
27
+ - gn
28
+ - ha
29
+ - ia
30
+ - id
31
+ - it
32
+ - ka
33
+ - ky
34
+ - lt
35
+ - lv
36
+ - mn
37
+ - mt
38
+ - nl
39
+ - or
40
+ - pl
41
+ - pt
42
+ - rm-sursilv
43
+ - rm-vallader
44
+ - ro
45
+ - ru
46
+ - rw
47
+ - sah
48
+ - sk
49
+ - sl
50
+ - sv-SE
51
+ - ta
52
+ - tr
53
+ - tt
54
+ - uk
55
+ - vi
56
+ - zh-CN
57
+ licenses:
58
+ - cc-by-4.0
59
+ multilinguality:
60
+ - multilingual
61
+ pretty_name: Multilingual Spoken Words
62
+ size_categories:
63
+ - 10M<n<100M
64
+ source_datasets:
65
+ - extended|common_voice
66
+ task_categories:
67
+ - speech-processing
68
+ task_ids:
69
+ - other-other-keyword-spotting
70
+ ---
71
+
72
+ # Dataset Card for Multilingual Spoken Words
73
+
74
+ ## Table of Contents
75
+ - [Table of Contents](#table-of-contents)
76
+ - [Dataset Description](#dataset-description)
77
+ - [Dataset Summary](#dataset-summary)
78
+ - [Supported Tasks and Leaderboards](#supported-tasks-and-leaderboards)
79
+ - [Languages](#languages)
80
+ - [Dataset Structure](#dataset-structure)
81
+ - [Data Instances](#data-instances)
82
+ - [Data Fields](#data-fields)
83
+ - [Data Splits](#data-splits)
84
+ - [Dataset Creation](#dataset-creation)
85
+ - [Curation Rationale](#curation-rationale)
86
+ - [Source Data](#source-data)
87
+ - [Annotations](#annotations)
88
+ - [Personal and Sensitive Information](#personal-and-sensitive-information)
89
+ - [Considerations for Using the Data](#considerations-for-using-the-data)
90
+ - [Social Impact of Dataset](#social-impact-of-dataset)
91
+ - [Discussion of Biases](#discussion-of-biases)
92
+ - [Other Known Limitations](#other-known-limitations)
93
+ - [Additional Information](#additional-information)
94
+ - [Dataset Curators](#dataset-curators)
95
+ - [Licensing Information](#licensing-information)
96
+ - [Citation Information](#citation-information)
97
+ - [Contributions](#contributions)
98
+
99
+ ## Dataset Description
100
+
101
+ - **Homepage:** https://mlcommons.org/en/multilingual-spoken-words/
102
+ - **Repository:** https://github.com/harvard-edge/multilingual_kws
103
+ - **Paper:** https://datasets-benchmarks-proceedings.neurips.cc/paper/2021/file/fe131d7f5a6b38b23cc967316c13dae2-Paper-round2.pdf
104
+ - **Leaderboard:**
105
+ - **Point of Contact:**
106
+
107
+ ### Dataset Summary
108
+
109
+ Multilingual Spoken Words Corpus is a large and growing audio dataset of spoken
110
+ words in 50 languages collectively spoken by over 5 billion people, for academic
111
+ research and commercial applications in keyword spotting and spoken term search,
112
+ licensed under CC-BY 4.0. The dataset contains more than 340,000 keywords,
113
+ totaling 23.4 million 1-second spoken examples (over 6,000 hours). The dataset
114
+ has many use cases, ranging from voice-enabled consumer devices to call center
115
+ automation. This dataset is generated by applying forced alignment on crowd-sourced sentence-level
116
+ audio to produce per-word timing estimates for extraction.
117
+ All alignments are included in the dataset.
118
+
119
+ ### Supported Tasks and Leaderboards
120
+
121
+ Keyword spotting, Spoken term search
122
+
123
+ ### Languages
124
+
125
+ The dataset is multilingual. To specify several languages to downloading pass a list of them to the
126
+ `language` argument:
127
+
128
+ ```python
129
+ ds = load_dataset("datasets/ml_spoken_words", languages=["ar", "tt", "br"])
130
+ ```
131
+
132
+ The dataset contains data for the following languages:
133
+
134
+ Low-resourced (<10 hours):
135
+ * Arabic (0.1G, 7.6h)
136
+ * Assamese (0.9M, 0.1h)
137
+ * Breton (69M, 5.6h)
138
+ * Chuvash (28M, 2.1h)
139
+ * Chinese (zh-CN) (42M, 3.1h)
140
+ * Dhivehi (0.7M, 0.04h)
141
+ * Frisian (0.1G, 9.6h)
142
+ * Georgian (20M, 1.4h)
143
+ * Guarani (0.7M, 1.3h)
144
+ * Greek (84M, 6.7h)
145
+ * Hakha Chin (26M, 0.1h)
146
+ * Hausa (90M, 1.0h)
147
+ * Interlingua (58M, 4.0h)
148
+ * Irish (38M, 3.2h)
149
+ * Latvian (51M, 4.2h)
150
+ * Lithuanian (21M, 0.46h)
151
+ * Maltese (88M, 7.3h)
152
+ * Oriya (0.7M, 0.1h)
153
+ * Romanian (59M, 4.5h)
154
+ * Sakha (42M, 3.3h)
155
+ * Slovenian (43M, 3.0h)
156
+ * Slovak (31M, 1.9h)
157
+ * Sursilvan (61M, 4.8h)
158
+ * Tamil (8.8M, 0.6h)
159
+ * Vallader (14M, 1.2h)
160
+ * Vietnamese (1.2M, 0.1h)
161
+
162
+ Medium-resourced (>10 & <100 hours):
163
+ * Czech (0.3G, 24h)
164
+ * Dutch (0.8G, 70h)
165
+ * Estonian (0.2G, 19h)
166
+ * Esperanto (1.3G, 77h)
167
+ * Indonesian (0.1G, 11h)
168
+ * Kyrgyz (0.1G, 12h)
169
+ * Mongolian (0.1G, 12h)
170
+ * Portuguese (0.7G, 58h)
171
+ * Swedish (0.1G, 12h)
172
+ * Tatar (4G, 30h)
173
+ * Turkish (1.3G, 29h)
174
+ * Ukrainian (0.2G, 18h)
175
+
176
+ Hig-resourced (>100 hours):
177
+ * Basque (1.7G, 118h)
178
+ * Catalan (8.7G, 615h)
179
+ * English (26G, 1957h)
180
+ * French (9.3G, 754h)
181
+ * German (14G, 1083h)
182
+ * Italian (2.2G, 155h)
183
+ * Kinyarwanda (6.1G, 422h)
184
+ * Persian (4.5G, 327h)
185
+ * Polish (1.8G, 130h)
186
+ * Russian (2.1G, 137h)
187
+ * Spanish (4.9G, 349h)
188
+ * Welsh (4.5G, 108h)
189
+
190
+ ## Dataset Structure
191
+
192
+ ### Data Instances
193
+
194
+ ```python
195
+ {'file': 'абзар_common_voice_tt_17737010.opus',
196
+ 'is_valid': True,
197
+ 'language': 0,
198
+ 'speaker_id': '687025afd5ce033048472754c8d2cb1cf8a617e469866bbdb3746e2bb2194202094a715906f91feb1c546893a5d835347f4869e7def2e360ace6616fb4340e38',
199
+ 'gender': 0,
200
+ 'keyword': 'абзар',
201
+ 'audio': {'path': 'абзар_common_voice_tt_17737010.opus',
202
+ 'array': array([2.03458695e-34, 2.03458695e-34, 2.03458695e-34, ...,
203
+ 2.03458695e-34, 2.03458695e-34, 2.03458695e-34]),
204
+ 'sampling_rate': 48000}}
205
+ ```
206
+
207
+ ### Data Fields
208
+
209
+ * file: strinrelative audio path inside the archive **#TODO: change according to the new local path schema?**
210
+ * is_valid: if a sample is valid
211
+ * language: language of an instance. Makes sense only when providing multiple languages to the
212
+ dataset loader (for example, `load_dataset("ml_spoken_words", languages=["ar", "tt"])`)
213
+ * speaker_id: unique id of a speaker. Can be "NA" if an instance is invalid
214
+ * gender: speaker gender. Can be one of `["MALE", "FEMALE", "OTHER", "NAN"]`
215
+ * keyword: word spoken in a current sample
216
+ * audio: a dictionary containing the relative path to the audio file,
217
+ the decoded audio array, and the sampling rate.
218
+ Note that when accessing the audio column: `dataset[0]["audio"]` the audio file is automatically
219
+ decoded and resampled to `dataset.features["audio"].sampling_rate`. Decoding and resampling of
220
+ a large number of audio files might take a significant amount of time.
221
+ Thus, it is important to first query the sample index before the "audio" column,
222
+ i.e. `dataset[0]["audio"]` should always be preferred over `dataset["audio"][0]`
223
+
224
+ ### Data Splits
225
+
226
+ The data for each language is splitted into train / validation / test parts.
227
+
228
+ ## Dataset Creation
229
+
230
+ ### Curation Rationale
231
+
232
+ [More Information Needed]
233
+
234
+ ### Source Data
235
+
236
+ #### Initial Data Collection and Normalization
237
+
238
+ The data comes form Common Voice dataset.
239
+
240
+ #### Who are the source language producers?
241
+
242
+ [More Information Needed]
243
+
244
+ ### Annotations
245
+
246
+ #### Annotation process
247
+
248
+ [More Information Needed]
249
+
250
+ #### Who are the annotators?
251
+
252
+ [More Information Needed]
253
+
254
+ ### Personal and Sensitive Information
255
+
256
+ he dataset consists of people who have donated their voice online.
257
+ You agree to not attempt to determine the identity of speakers.
258
+
259
+ ## Considerations for Using the Data
260
+
261
+ ### Social Impact of Dataset
262
+
263
+ [More Information Needed]
264
+
265
+ ### Discussion of Biases
266
+
267
+ [More Information Needed]
268
+
269
+ ### Other Known Limitations
270
+
271
+ [More Information Needed]
272
+
273
+ ## Additional Information
274
+
275
+ ### Dataset Curators
276
+
277
+ [More Information Needed]
278
+
279
+ ### Licensing Information
280
+
281
+ The dataset is licensed under [CC-BY 4.0](https://creativecommons.org/licenses/by/4.0/) and can be used for academic
282
+ research and commercial applications in keyword spotting and spoken term search.
283
+
284
+ ### Citation Information
285
+
286
+ ```
287
+ @inproceedings{mazumder2021multilingual,
288
+ title={Multilingual Spoken Words Corpus},
289
+ author={Mazumder, Mark and Chitlangia, Sharad and Banbury, Colby and Kang, Yiping and Ciro, Juan Manuel and Achorn, Keith and Galvez, Daniel and Sabini, Mark and Mattson, Peter and Kanter, David and others},
290
+ booktitle={Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2)},
291
+ year={2021}
292
+ }
293
+ ```
294
+
295
+ ### Contributions
296
+
297
+ Thanks to [@polinaeterna](https://github.com/polinaeterna) for adding this dataset.
dataset_infos.json ADDED
The diff for this file is too large to render. See raw diff
 
ml_spoken_words.py ADDED
@@ -0,0 +1,244 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # coding=utf-8
2
+ # Copyright 2022 The HuggingFace Datasets Authors and the current dataset script contributor.
3
+ #
4
+ # Licensed under the Apache License, Version 2.0 (the "License");
5
+ # you may not use this file except in compliance with the License.
6
+ # You may obtain a copy of the License at
7
+ #
8
+ # http://www.apache.org/licenses/LICENSE-2.0
9
+ #
10
+ # Unless required by applicable law or agreed to in writing, software
11
+ # distributed under the License is distributed on an "AS IS" BASIS,
12
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13
+ # See the License for the specific language governing permissions and
14
+ # limitations under the License.
15
+
16
+ """
17
+ Multilingual Spoken Words Corpus is a large and growing audio dataset of spoken
18
+ words in 50 languages collectively spoken by over 5 billion people, for academic
19
+ research and commercial applications in keyword spotting and spoken term search,
20
+ licensed under CC-BY 4.0. The dataset contains more than 340,000 keywords,
21
+ totaling 23.4 million 1-second spoken examples (over 6,000 hours).
22
+ """
23
+
24
+
25
+ import csv
26
+ from functools import partial
27
+
28
+ import datasets
29
+
30
+
31
+ _CITATION = """\
32
+ @inproceedings{mazumder2021multilingual,
33
+ title={Multilingual Spoken Words Corpus},
34
+ author={Mazumder, Mark and Chitlangia, Sharad and Banbury, Colby and Kang, Yiping and Ciro, Juan Manuel and Achorn, Keith and Galvez, Daniel and Sabini, Mark and Mattson, Peter and Kanter, David and others},
35
+ booktitle={Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2)},
36
+ year={2021}
37
+ }
38
+ """
39
+
40
+ _DESCRIPTION = """\
41
+ Multilingual Spoken Words Corpus is a large and growing audio dataset of spoken
42
+ words in 50 languages collectively spoken by over 5 billion people, for academic
43
+ research and commercial applications in keyword spotting and spoken term search,
44
+ licensed under CC-BY 4.0. The dataset contains more than 340,000 keywords,
45
+ totaling 23.4 million 1-second spoken examples (over 6,000 hours). The dataset
46
+ has many use cases, ranging from voice-enabled consumer devices to call center
47
+ automation. This dataset is generated by applying forced alignment on crowd-sourced sentence-level
48
+ audio to produce per-word timing estimates for extraction.
49
+ All alignments are included in the dataset.
50
+ """
51
+
52
+ _HOMEPAGE = "https://mlcommons.org/en/multilingual-spoken-words/"
53
+
54
+ _LICENSE = "CC-BY 4.0."
55
+
56
+ _VERSION = datasets.Version("1.0.0")
57
+
58
+ _BASE_URL = "https://huggingface.co/datasets/polinaeterna/ml_spoken_words/resolve/main/data/{lang}/"
59
+ _AUDIO_URL = _BASE_URL + "{split}/audio/{n}.tar.gz"
60
+ _SPLITS_URL = _BASE_URL + "splits.tar.gz"
61
+ _N_FILES_URL = _BASE_URL + "{split}/n_files.txt"
62
+
63
+ _GENDERS = ["MALE", "FEMALE", "OTHER", "NAN"]
64
+
65
+ _LANGUAGES = [
66
+ "ar",
67
+ "as",
68
+ "br",
69
+ "ca",
70
+ "cnh",
71
+ "cs",
72
+ "cv",
73
+ "cy",
74
+ "de",
75
+ "dv",
76
+ "el",
77
+ "en",
78
+ "eo",
79
+ "es",
80
+ "et",
81
+ "eu",
82
+ "fa",
83
+ "fr",
84
+ "fy-NL",
85
+ "ga-IE",
86
+ "gn",
87
+ "ha",
88
+ "ia",
89
+ "id",
90
+ "it",
91
+ "ka",
92
+ "ky",
93
+ "lt",
94
+ "lv",
95
+ "mn",
96
+ "mt",
97
+ "nl",
98
+ "or",
99
+ "pl",
100
+ "pt",
101
+ "rm-sursilv",
102
+ "rm-vallader",
103
+ "ro",
104
+ "ru",
105
+ "rw",
106
+ "sah",
107
+ "sk",
108
+ "sl",
109
+ "sv-SE",
110
+ "ta",
111
+ "tr",
112
+ "tt",
113
+ "uk",
114
+ "vi",
115
+ "zh-CN",
116
+ ]
117
+
118
+
119
+ class MlSpokenWordsConfig(datasets.BuilderConfig):
120
+ """BuilderConfig for MlSpokenWords."""
121
+
122
+ def __init__(self, *args, languages, **kwargs):
123
+ """BuilderConfig for MlSpokenWords.
124
+ Args:
125
+ languages (:obj:`Union[List[str], str]`): language or list of languages to load
126
+ **kwargs: keyword arguments forwarded to super.
127
+ """
128
+ super().__init__(
129
+ *args,
130
+ name="+".join(languages) if isinstance(languages, list) else languages,
131
+ **kwargs,
132
+ )
133
+ self.languages = languages if isinstance(languages, list) else [languages]
134
+
135
+
136
+ class MlSpokenWords(datasets.GeneratorBasedBuilder):
137
+ """
138
+ Multilingual Spoken Words Corpus is a large and growing audio dataset of spoken
139
+ words in 50 languages collectively spoken by over 5 billion people, for academic
140
+ research and commercial applications in keyword spotting and spoken term search,
141
+ licensed under CC-BY 4.0. The dataset contains more than 340,000 keywords,
142
+ totaling 23.4 million 1-second spoken examples (over 6,000 hours).
143
+ """
144
+
145
+ VERSION = _VERSION
146
+ BUILDER_CONFIGS = [MlSpokenWordsConfig(languages=[lang], version=_VERSION) for lang in _LANGUAGES]
147
+ BUILDER_CONFIG_CLASS = MlSpokenWordsConfig
148
+
149
+ def _info(self):
150
+ features = datasets.Features(
151
+ {
152
+ "file": datasets.Value("string"),
153
+ "is_valid": datasets.Value("bool"),
154
+ "language": datasets.ClassLabel(names=self.config.languages),
155
+ "speaker_id": datasets.Value("string"),
156
+ "gender": datasets.ClassLabel(names=_GENDERS),
157
+ "keyword": datasets.Value("string"), # seems that there are too many of them (340k unique keywords)
158
+ "audio": datasets.Audio(sampling_rate=48_000),
159
+ }
160
+ )
161
+ return datasets.DatasetInfo(
162
+ description=_DESCRIPTION,
163
+ features=features,
164
+ homepage=_HOMEPAGE,
165
+ license=_LICENSE,
166
+ citation=_CITATION,
167
+ )
168
+
169
+ def _split_generators(self, dl_manager):
170
+ splits_archive_path = [dl_manager.download(_SPLITS_URL.format(lang=lang)) for lang in self.config.languages]
171
+ download_audio = partial(_download_audio_archives, dl_manager=dl_manager)
172
+
173
+ return [
174
+ datasets.SplitGenerator(
175
+ name=datasets.Split.TRAIN,
176
+ gen_kwargs={
177
+ "audio_archives": [download_audio(split="train", lang=lang) for lang in self.config.languages],
178
+ "splits_archives": [dl_manager.iter_archive(path) for path in splits_archive_path],
179
+ "split": "train",
180
+ },
181
+ ),
182
+ datasets.SplitGenerator(
183
+ name=datasets.Split.VALIDATION,
184
+ gen_kwargs={
185
+ "audio_archives": [download_audio(split="dev", lang=lang) for lang in self.config.languages],
186
+ "splits_archives": [dl_manager.iter_archive(path) for path in splits_archive_path],
187
+ "split": "dev",
188
+ },
189
+ ),
190
+ datasets.SplitGenerator(
191
+ name=datasets.Split.TEST,
192
+ gen_kwargs={
193
+ "audio_archives": [download_audio(split="test", lang=lang) for lang in self.config.languages],
194
+ "splits_archives": [dl_manager.iter_archive(path) for path in splits_archive_path],
195
+ "split": "test",
196
+ },
197
+ ),
198
+ ]
199
+
200
+ def _generate_examples(self, audio_archives, splits_archives, split):
201
+ metadata = dict()
202
+ for lang_idx, lang in enumerate(self.config.languages):
203
+ for split_filename, split_file in splits_archives[lang_idx]:
204
+ if split_filename.split(".csv")[0] == split:
205
+ csv_reader = csv.reader([line.decode("utf-8") for line in split_file.readlines()], delimiter=",")
206
+ for i, (link, word, is_valid, speaker, gender) in enumerate(csv_reader):
207
+ if i == 0:
208
+ continue
209
+ audio_filename = "_".join(link.split("/"))
210
+ metadata[audio_filename] = {
211
+ "keyword": word,
212
+ "is_valid": is_valid,
213
+ "speaker_id": speaker,
214
+ "gender": gender if gender and gender != "NA" else "NAN", # some values are "NA"
215
+ }
216
+
217
+ for audio_archive in audio_archives[lang_idx]:
218
+ for audio_filename, audio_file in audio_archive:
219
+ yield audio_filename, {
220
+ "file": audio_filename,
221
+ "language": lang,
222
+ "audio": {"path": audio_filename, "bytes": audio_file.read()},
223
+ **metadata[audio_filename],
224
+ }
225
+
226
+
227
+ def _download_audio_archives(dl_manager, lang, split):
228
+ """
229
+ All audio files are stored in several .tar.gz archives with names like 0.tar.gz, 1.tar.gz, ...
230
+ Number of archives stored in a separate .txt file (n_files.txt)
231
+
232
+ Prepare all the audio archives for iterating over them and their audio files.
233
+ """
234
+
235
+ n_files_url = _N_FILES_URL.format(lang=lang, split=split)
236
+ n_files_path = dl_manager.download(n_files_url)
237
+
238
+ with open(n_files_path, "r", encoding="utf-8") as file:
239
+ n_files = int(file.read().strip()) # the file contains a number of archives
240
+
241
+ archive_urls = [_AUDIO_URL.format(lang=lang, split=split, n=i) for i in range(n_files)]
242
+ archive_paths = dl_manager.download(archive_urls)
243
+
244
+ return [dl_manager.iter_archive(archive_path) for archive_path in archive_paths]