system HF staff commited on
Commit
cb0e6fc
1 Parent(s): 64cab28

Update files from the datasets library (from 1.12.0)

Browse files

Release notes: https://github.com/huggingface/datasets/releases/tag/1.12.0

README.md CHANGED
@@ -15,6 +15,8 @@ size_categories:
15
  source_datasets:
16
  - original
17
  - extended|librispeech_asr
 
 
18
  task_categories:
19
  - speech-processing
20
  task_ids:
@@ -106,7 +108,46 @@ Automatic Speaker Verification (ASV) verifies whether the speakers of a pair of
106
 
107
  #### sd
108
 
109
- Speaker Diarization (SD) predicts who is speaking when for each timestamp, and multiple speakers can speak simultaneously. The model has to encode rich speaker characteristics for each frame and should be able to represent mixtures of signals. [LibriMix](https://github.com/s3prl/s3prl/tree/master/downstream#sd-speaker-diarization) is adopted where LibriSpeech train-clean-100/dev-clean/test-clean are used to generate mixtures for training/validation/testing. We focus on the two-speaker scenario as the first step. The time-coded speaker labels were generated using alignments from Kaldi LibriSpeech ASR model. The evaluation metric is diarization error rate (DER).
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
110
 
111
  #### er
112
 
@@ -130,7 +171,7 @@ The language data in SUPERB is in English (BCP-47 `en`)
130
 
131
  An example from each split looks like:
132
 
133
- ```json
134
  {'chapter_id': 1240,
135
  'file': 'path/to/file.flac',
136
  'id': '103-1240-0000',
@@ -143,8 +184,14 @@ An example from each split looks like:
143
 
144
  #### ks
145
 
146
- [More Information Needed](https://github.com/huggingface/datasets/blob/master/CONTRIBUTING.md#how-to-contribute-to-the-dataset-cards)
147
 
 
 
 
 
 
 
148
 
149
  #### qbe
150
 
@@ -173,7 +220,19 @@ An example from each split looks like:
173
 
174
  #### sd
175
 
176
- [More Information Needed](https://github.com/huggingface/datasets/blob/master/CONTRIBUTING.md#how-to-contribute-to-the-dataset-cards)
 
 
 
 
 
 
 
 
 
 
 
 
177
 
178
 
179
  #### er
@@ -194,14 +253,14 @@ An example from each split looks like:
194
 
195
  - `file`: a `string` feature.
196
  - `text`: a `string` feature.
197
- - `speaker_id`: a `int64` feature
198
- - `chapter_id`: a `int64` feature
199
- - `id`: a `string` feature
200
 
201
  #### ks
202
 
203
- [More Information Needed](https://github.com/huggingface/datasets/blob/master/CONTRIBUTING.md#how-to-contribute-to-the-dataset-cards)
204
-
205
 
206
  #### qbe
207
 
@@ -230,8 +289,15 @@ An example from each split looks like:
230
 
231
  #### sd
232
 
233
- [More Information Needed](https://github.com/huggingface/datasets/blob/master/CONTRIBUTING.md#how-to-contribute-to-the-dataset-cards)
234
-
 
 
 
 
 
 
 
235
 
236
  #### er
237
 
@@ -252,8 +318,9 @@ An example from each split looks like:
252
 
253
  #### ks
254
 
255
- [More Information Needed](https://github.com/huggingface/datasets/blob/master/CONTRIBUTING.md#how-to-contribute-to-the-dataset-cards)
256
-
 
257
 
258
  #### qbe
259
 
@@ -282,8 +349,11 @@ An example from each split looks like:
282
 
283
  #### sd
284
 
285
- [More Information Needed](https://github.com/huggingface/datasets/blob/master/CONTRIBUTING.md#how-to-contribute-to-the-dataset-cards)
286
 
 
 
 
287
 
288
  #### er
289
 
@@ -385,4 +455,4 @@ the correct citation for each contained dataset.
385
 
386
  ### Contributions
387
 
388
- Thanks to [@lewtun](https://github.com/lewtun) and [@albertvillanova](https://github.com/albertvillanova) for adding this dataset.
 
15
  source_datasets:
16
  - original
17
  - extended|librispeech_asr
18
+ - extended|other-librimix
19
+ - extended|other-speech_commands
20
  task_categories:
21
  - speech-processing
22
  task_ids:
 
108
 
109
  #### sd
110
 
111
+ Speaker Diarization (SD) predicts *who is speaking when* for each timestamp, and multiple speakers can speak simultaneously. The model has to encode rich speaker characteristics for each frame and should be able to represent mixtures of signals. [LibriMix](https://github.com/s3prl/s3prl/tree/master/downstream#sd-speaker-diarization) is adopted where LibriSpeech train-clean-100/dev-clean/test-clean are used to generate mixtures for training/validation/testing. We focus on the two-speaker scenario as the first step. The time-coded speaker labels were generated using alignments from Kaldi LibriSpeech ASR model. The evaluation metric is diarization error rate (DER).
112
+
113
+ ##### Example of usage
114
+
115
+ Use these auxiliary functions to:
116
+ - load the audio file into an audio data array
117
+ - generate the label array
118
+
119
+ ```python
120
+ def load_audio_file(example, frame_shift=160):
121
+ import soundfile as sf
122
+
123
+ example["array"], example["sample_rate"] = sf.read(
124
+ example["file"], start=example["start"] * frame_shift, stop=example["end"] * frame_shift
125
+ )
126
+ return example
127
+
128
+
129
+ def generate_label(example, frame_shift=160, num_speakers=2, rate=16000):
130
+ import numpy as np
131
+
132
+ start = example["start"]
133
+ end = example["end"]
134
+ frame_num = end - start
135
+ speakers = sorted({speaker["speaker_id"] for speaker in example["speakers"]})
136
+ label = np.zeros((frame_num, num_speakers), dtype=np.int32)
137
+ for speaker in example["speakers"]:
138
+ speaker_index = speakers.index(speaker["speaker_id"])
139
+ start_frame = np.rint(speaker["start"] * rate / frame_shift).astype(int)
140
+ end_frame = np.rint(speaker["end"] * rate / frame_shift).astype(int)
141
+ rel_start = rel_end = None
142
+ if start <= start_frame < end:
143
+ rel_start = start_frame - start
144
+ if start < end_frame <= end:
145
+ rel_end = end_frame - start
146
+ if rel_start is not None or rel_end is not None:
147
+ label[rel_start:rel_end, speaker_index] = 1
148
+ example["label"] = label
149
+ return example
150
+ ```
151
 
152
  #### er
153
 
 
171
 
172
  An example from each split looks like:
173
 
174
+ ```python
175
  {'chapter_id': 1240,
176
  'file': 'path/to/file.flac',
177
  'id': '103-1240-0000',
 
184
 
185
  #### ks
186
 
187
+ An example from each split looks like:
188
 
189
+ ```python
190
+ {
191
+ 'file': '/path/yes/af7a8296_nohash_1.wav',
192
+ 'label': 'yes'
193
+ }
194
+ ```
195
 
196
  #### qbe
197
 
 
220
 
221
  #### sd
222
 
223
+ An example from each split looks like:
224
+ ```python
225
+ {
226
+ 'record_id': '1578-6379-0038_6415-111615-0009',
227
+ 'file': 'path/to/file.wav',
228
+ 'start': 0,
229
+ 'end': 1590,
230
+ 'speakers': [
231
+ {'speaker_id': '1578', 'start': 28, 'end': 657},
232
+ {'speaker_id': '6415', 'start': 28, 'end': 1576}
233
+ ]
234
+ }
235
+ ```
236
 
237
 
238
  #### er
 
253
 
254
  - `file`: a `string` feature.
255
  - `text`: a `string` feature.
256
+ - `speaker_id`: a `int64` feature.
257
+ - `chapter_id`: a `int64` feature.
258
+ - `id`: a `string` feature.
259
 
260
  #### ks
261
 
262
+ - `file` (`string`): Path to the WAV audio file.
263
+ - `label` (`string`): Label of the spoken command.
264
 
265
  #### qbe
266
 
 
289
 
290
  #### sd
291
 
292
+ The data fields in all splits are:
293
+ - `record_id` (`string`): ID of the record.
294
+ - `file` (`string`): Path to the WAV audio file.
295
+ - `start` (`integer`): Start frame of the audio.
296
+ - `end` (`integer`): End frame of the audio.
297
+ - `speakers` (`list` of `dict`): List of speakers in the audio. Each item contains the fields:
298
+ - `speaker_id` (`string`): ID of the speaker.
299
+ - `start` (`integer`): Frame when the speaker starts speaking.
300
+ - `end` (`integer`): Frame when the speaker stops speaking.
301
 
302
  #### er
303
 
 
318
 
319
  #### ks
320
 
321
+ | | train | validation | test |
322
+ |----|------:|-----------:|-----:|
323
+ | ks | 51094 | 6798 | 3081 |
324
 
325
  #### qbe
326
 
 
349
 
350
  #### sd
351
 
352
+ The data is split into "train", "dev" and "test" sets, each containing the following number of examples:
353
 
354
+ | | train | dev | test |
355
+ |----|------:|-----:|-----:|
356
+ | sd | 13901 | 3014 | 3002 |
357
 
358
  #### er
359
 
 
455
 
456
  ### Contributions
457
 
458
+ Thanks to [@lewtun](https://github.com/lewtun), [@albertvillanova](https://github.com/albertvillanova) and [@anton-l](https://github.com/anton-l) for adding this dataset.
dataset_infos.json CHANGED
@@ -1 +1 @@
1
- {"asr": {"description": "Self-supervised learning (SSL) has proven vital for advancing research in\nnatural language processing (NLP) and computer vision (CV). The paradigm\npretrains a shared model on large volumes of unlabeled data and achieves\nstate-of-the-art (SOTA) for various tasks with minimal adaptation. However, the\nspeech processing community lacks a similar setup to systematically explore the\nparadigm. To bridge this gap, we introduce Speech processing Universal\nPERformance Benchmark (SUPERB). SUPERB is a leaderboard to benchmark the\nperformance of a shared model across a wide range of speech processing tasks\nwith minimal architecture changes and labeled data. Among multiple usages of the\nshared model, we especially focus on extracting the representation learned from\nSSL due to its preferable re-usability. We present a simple framework to solve\nSUPERB tasks by learning task-specialized lightweight prediction heads on top of\nthe frozen shared model. Our results demonstrate that the framework is promising\nas SSL representations show competitive generalizability and accessibility\nacross SUPERB tasks. We release SUPERB as a challenge with a leaderboard and a\nbenchmark toolkit to fuel the research in representation learning and general\nspeech processing.\n\nNote that in order to limit the required storage for preparing this dataset, the\naudio is stored in the .flac format and is not converted to a float32 array. To\nconvert, the audio file to a float32 array, please make use of the `.map()`\nfunction as follows:\n\n\n```python\nimport soundfile as sf\n\ndef map_to_array(batch):\n speech_array, _ = sf.read(batch[\"file\"])\n batch[\"speech\"] = speech_array\n return batch\n\ndataset = dataset.map(map_to_array, remove_columns=[\"file\"])\n```\n", "citation": "@article{DBLP:journals/corr/abs-2105-01051,\n author = {Shu{-}Wen Yang and\n Po{-}Han Chi and\n Yung{-}Sung Chuang and\n Cheng{-}I Jeff Lai and\n Kushal Lakhotia and\n Yist Y. Lin and\n Andy T. Liu and\n Jiatong Shi and\n Xuankai Chang and\n Guan{-}Ting Lin and\n Tzu{-}Hsien Huang and\n Wei{-}Cheng Tseng and\n Ko{-}tik Lee and\n Da{-}Rong Liu and\n Zili Huang and\n Shuyan Dong and\n Shang{-}Wen Li and\n Shinji Watanabe and\n Abdelrahman Mohamed and\n Hung{-}yi Lee},\n title = {{SUPERB:} Speech processing Universal PERformance Benchmark},\n journal = {CoRR},\n volume = {abs/2105.01051},\n year = {2021},\n url = {https://arxiv.org/abs/2105.01051},\n archivePrefix = {arXiv},\n eprint = {2105.01051},\n timestamp = {Thu, 01 Jul 2021 13:30:22 +0200},\n biburl = {https://dblp.org/rec/journals/corr/abs-2105-01051.bib},\n bibsource = {dblp computer science bibliography, https://dblp.org}\n}\n", "homepage": "http://www.openslr.org/12", "license": "", "features": {"file": {"dtype": "string", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}, "speaker_id": {"dtype": "int64", "id": null, "_type": "Value"}, "chapter_id": {"dtype": "int64", "id": null, "_type": "Value"}, "id": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": {"input": "file", "output": "text"}, "task_templates": [{"task": "automatic-speech-recognition", "audio_file_path_column": "file", "transcription_column": "text"}], "builder_name": "superb", "config_name": "asr", "version": {"version_str": "1.9.0", "description": "", "major": 1, "minor": 9, "patch": 0}, "splits": {"train": {"name": "train", "num_bytes": 11823891, "num_examples": 28539, "dataset_name": "superb"}, "validation": {"name": "validation", "num_bytes": 894510, "num_examples": 2703, "dataset_name": "superb"}, "test": {"name": "test", "num_bytes": 868614, "num_examples": 2620, "dataset_name": "superb"}}, "download_checksums": {"http://www.openslr.org/resources/12/dev-clean.tar.gz": {"num_bytes": 337926286, "checksum": "76f87d090650617fca0cac8f88b9416e0ebf80350acb97b343a85fa903728ab3"}, "http://www.openslr.org/resources/12/test-clean.tar.gz": {"num_bytes": 346663984, "checksum": "39fde525e59672dc6d1551919b1478f724438a95aa55f874b576be21967e6c23"}, "http://www.openslr.org/resources/12/train-clean-100.tar.gz": {"num_bytes": 6387309499, "checksum": "d4ddd1d5a6ab303066f14971d768ee43278a5f2a0aa43dc716b0e64ecbbbf6e2"}}, "download_size": 7071899769, "post_processing_size": null, "dataset_size": 13587015, "size_in_bytes": 7085486784}}
 
1
+ {"asr": {"description": "Self-supervised learning (SSL) has proven vital for advancing research in\nnatural language processing (NLP) and computer vision (CV). The paradigm\npretrains a shared model on large volumes of unlabeled data and achieves\nstate-of-the-art (SOTA) for various tasks with minimal adaptation. However, the\nspeech processing community lacks a similar setup to systematically explore the\nparadigm. To bridge this gap, we introduce Speech processing Universal\nPERformance Benchmark (SUPERB). SUPERB is a leaderboard to benchmark the\nperformance of a shared model across a wide range of speech processing tasks\nwith minimal architecture changes and labeled data. Among multiple usages of the\nshared model, we especially focus on extracting the representation learned from\nSSL due to its preferable re-usability. We present a simple framework to solve\nSUPERB tasks by learning task-specialized lightweight prediction heads on top of\nthe frozen shared model. Our results demonstrate that the framework is promising\nas SSL representations show competitive generalizability and accessibility\nacross SUPERB tasks. We release SUPERB as a challenge with a leaderboard and a\nbenchmark toolkit to fuel the research in representation learning and general\nspeech processing.\n\nNote that in order to limit the required storage for preparing this dataset, the\naudio is stored in the .flac format and is not converted to a float32 array. To\nconvert, the audio file to a float32 array, please make use of the `.map()`\nfunction as follows:\n\n\n```python\nimport soundfile as sf\n\ndef map_to_array(batch):\n speech_array, _ = sf.read(batch[\"file\"])\n batch[\"speech\"] = speech_array\n return batch\n\ndataset = dataset.map(map_to_array, remove_columns=[\"file\"])\n```\n", "citation": "@article{DBLP:journals/corr/abs-2105-01051,\n author = {Shu{-}Wen Yang and\n Po{-}Han Chi and\n Yung{-}Sung Chuang and\n Cheng{-}I Jeff Lai and\n Kushal Lakhotia and\n Yist Y. Lin and\n Andy T. Liu and\n Jiatong Shi and\n Xuankai Chang and\n Guan{-}Ting Lin and\n Tzu{-}Hsien Huang and\n Wei{-}Cheng Tseng and\n Ko{-}tik Lee and\n Da{-}Rong Liu and\n Zili Huang and\n Shuyan Dong and\n Shang{-}Wen Li and\n Shinji Watanabe and\n Abdelrahman Mohamed and\n Hung{-}yi Lee},\n title = {{SUPERB:} Speech processing Universal PERformance Benchmark},\n journal = {CoRR},\n volume = {abs/2105.01051},\n year = {2021},\n url = {https://arxiv.org/abs/2105.01051},\n archivePrefix = {arXiv},\n eprint = {2105.01051},\n timestamp = {Thu, 01 Jul 2021 13:30:22 +0200},\n biburl = {https://dblp.org/rec/journals/corr/abs-2105-01051.bib},\n bibsource = {dblp computer science bibliography, https://dblp.org}\n}\n", "homepage": "http://www.openslr.org/12", "license": "", "features": {"file": {"dtype": "string", "id": null, "_type": "Value"}, "text": {"dtype": "string", "id": null, "_type": "Value"}, "speaker_id": {"dtype": "int64", "id": null, "_type": "Value"}, "chapter_id": {"dtype": "int64", "id": null, "_type": "Value"}, "id": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": {"input": "file", "output": "text"}, "task_templates": [{"task": "automatic-speech-recognition", "audio_file_path_column": "file", "transcription_column": "text"}], "builder_name": "superb", "config_name": "asr", "version": {"version_str": "1.9.0", "description": "", "major": 1, "minor": 9, "patch": 0}, "splits": {"train": {"name": "train", "num_bytes": 11852430, "num_examples": 28539, "dataset_name": "superb"}, "validation": {"name": "validation", "num_bytes": 897213, "num_examples": 2703, "dataset_name": "superb"}, "test": {"name": "test", "num_bytes": 871234, "num_examples": 2620, "dataset_name": "superb"}}, "download_checksums": {"http://www.openslr.org/resources/12/dev-clean.tar.gz": {"num_bytes": 337926286, "checksum": "76f87d090650617fca0cac8f88b9416e0ebf80350acb97b343a85fa903728ab3"}, "http://www.openslr.org/resources/12/test-clean.tar.gz": {"num_bytes": 346663984, "checksum": "39fde525e59672dc6d1551919b1478f724438a95aa55f874b576be21967e6c23"}, "http://www.openslr.org/resources/12/train-clean-100.tar.gz": {"num_bytes": 6387309499, "checksum": "d4ddd1d5a6ab303066f14971d768ee43278a5f2a0aa43dc716b0e64ecbbbf6e2"}}, "download_size": 7071899769, "post_processing_size": null, "dataset_size": 13620877, "size_in_bytes": 7085520646}, "sd": {"description": "Self-supervised learning (SSL) has proven vital for advancing research in\nnatural language processing (NLP) and computer vision (CV). The paradigm\npretrains a shared model on large volumes of unlabeled data and achieves\nstate-of-the-art (SOTA) for various tasks with minimal adaptation. However, the\nspeech processing community lacks a similar setup to systematically explore the\nparadigm. To bridge this gap, we introduce Speech processing Universal\nPERformance Benchmark (SUPERB). SUPERB is a leaderboard to benchmark the\nperformance of a shared model across a wide range of speech processing tasks\nwith minimal architecture changes and labeled data. Among multiple usages of the\nshared model, we especially focus on extracting the representation learned from\nSSL due to its preferable re-usability. We present a simple framework to solve\nSUPERB tasks by learning task-specialized lightweight prediction heads on top of\nthe frozen shared model. Our results demonstrate that the framework is promising\nas SSL representations show competitive generalizability and accessibility\nacross SUPERB tasks. We release SUPERB as a challenge with a leaderboard and a\nbenchmark toolkit to fuel the research in representation learning and general\nspeech processing.\n\nNote that in order to limit the required storage for preparing this dataset, the\naudio is stored in the .flac format and is not converted to a float32 array. To\nconvert, the audio file to a float32 array, please make use of the `.map()`\nfunction as follows:\n\n\n```python\nimport soundfile as sf\n\ndef map_to_array(batch):\n speech_array, _ = sf.read(batch[\"file\"])\n batch[\"speech\"] = speech_array\n return batch\n\ndataset = dataset.map(map_to_array, remove_columns=[\"file\"])\n```\n", "citation": "@article{DBLP:journals/corr/abs-2105-01051,\n author = {Shu{-}Wen Yang and\n Po{-}Han Chi and\n Yung{-}Sung Chuang and\n Cheng{-}I Jeff Lai and\n Kushal Lakhotia and\n Yist Y. Lin and\n Andy T. Liu and\n Jiatong Shi and\n Xuankai Chang and\n Guan{-}Ting Lin and\n Tzu{-}Hsien Huang and\n Wei{-}Cheng Tseng and\n Ko{-}tik Lee and\n Da{-}Rong Liu and\n Zili Huang and\n Shuyan Dong and\n Shang{-}Wen Li and\n Shinji Watanabe and\n Abdelrahman Mohamed and\n Hung{-}yi Lee},\n title = {{SUPERB:} Speech processing Universal PERformance Benchmark},\n journal = {CoRR},\n volume = {abs/2105.01051},\n year = {2021},\n url = {https://arxiv.org/abs/2105.01051},\n archivePrefix = {arXiv},\n eprint = {2105.01051},\n timestamp = {Thu, 01 Jul 2021 13:30:22 +0200},\n biburl = {https://dblp.org/rec/journals/corr/abs-2105-01051.bib},\n bibsource = {dblp computer science bibliography, https://dblp.org}\n}\n", "homepage": "https://github.com/ftshijt/LibriMix", "license": "", "features": {"record_id": {"dtype": "string", "id": null, "_type": "Value"}, "file": {"dtype": "string", "id": null, "_type": "Value"}, "start": {"dtype": "int64", "id": null, "_type": "Value"}, "end": {"dtype": "int64", "id": null, "_type": "Value"}, "speakers": [{"speaker_id": {"dtype": "string", "id": null, "_type": "Value"}, "start": {"dtype": "int64", "id": null, "_type": "Value"}, "end": {"dtype": "int64", "id": null, "_type": "Value"}}]}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "superb", "config_name": "sd", "version": {"version_str": "1.9.0", "description": "", "major": 1, "minor": 9, "patch": 0}, "splits": {"train": {"name": "train", "num_bytes": 4622013, "num_examples": 13901, "dataset_name": "superb"}, "dev": {"name": "dev", "num_bytes": 860472, "num_examples": 3014, "dataset_name": "superb"}, "test": {"name": "test", "num_bytes": 847803, "num_examples": 3002, "dataset_name": "superb"}}, "download_checksums": {"https://huggingface.co/datasets/superb/superb-data/resolve/main/sd/train/reco2dur": {"num_bytes": 540906, "checksum": "879dca4b1108c93bd86df879463fca15a4de42a0f95a7e6987138dc6029b5554"}, "https://huggingface.co/datasets/superb/superb-data/resolve/main/sd/train/segments": {"num_bytes": 5723993, "checksum": "f19cb0ecc342f8d2cd855118879a111822d7cf55fcd078ef156f5147233a8e11"}, "https://huggingface.co/datasets/superb/superb-data/resolve/main/sd/train/utt2spk": {"num_bytes": 3165995, "checksum": "a4295726caf05d72f5ad24706180b9dbccffe6c0c2fc0128ca4b02b7b828a28a"}, "https://huggingface.co/datasets/superb/superb-data/resolve/main/sd/train/wav.zip": {"num_bytes": 5706733518, "checksum": "4231070427ffbc9b3bddae874dba32f3985a0db0b0feb4dfa29ed4d1d11bf41b"}, "https://huggingface.co/datasets/superb/superb-data/resolve/main/sd/dev/reco2dur": {"num_bytes": 115918, "checksum": "a30fd59ad01db0315a82cad7a64baea009e6c2bcdfb6b2501bc8873ede72de06"}, "https://huggingface.co/datasets/superb/superb-data/resolve/main/sd/dev/segments": {"num_bytes": 673006, "checksum": "2b977917e7ab9feec03afb4fd6a4662df90e48dbcc42977a4b9c89c8d40432ee"}, "https://huggingface.co/datasets/superb/superb-data/resolve/main/sd/dev/utt2spk": {"num_bytes": 374794, "checksum": "9f47a7bed76e7a03e57d66ba9cc5f57d85d91f748d0b1eb20301d09e6c24cd20"}, "https://huggingface.co/datasets/superb/superb-data/resolve/main/sd/dev/wav.zip": {"num_bytes": 765594100, "checksum": "e28b3422ce59e2a5273be924e6ed6b8f115c0983db1997e56441973c27ee1cd8"}, "https://huggingface.co/datasets/superb/superb-data/resolve/main/sd/test/reco2dur": {"num_bytes": 113357, "checksum": "6e013d917015031e2f1383871b52dfc1122e7b16cdee53bd8e5e0a7fbc57e406"}, "https://huggingface.co/datasets/superb/superb-data/resolve/main/sd/test/segments": {"num_bytes": 650742, "checksum": "92f8de0f56c55a34e9111542c24ea13f2d2efaf9ebe64af31250cadab020f987"}, "https://huggingface.co/datasets/superb/superb-data/resolve/main/sd/test/utt2spk": {"num_bytes": 361548, "checksum": "19dcb558aa886f0d553d8d9b8735ea1998b83e96d5245e5511cb732c84625ffd"}, "https://huggingface.co/datasets/superb/superb-data/resolve/main/sd/test/wav.zip": {"num_bytes": 706322334, "checksum": "9c8ee97d3068759c0101bf88684abab77183374dbb3bb40f7c0b25d385992ea6"}}, "download_size": 7190370211, "post_processing_size": null, "dataset_size": 6330288, "size_in_bytes": 7196700499}, "ks": {"description": "Self-supervised learning (SSL) has proven vital for advancing research in\nnatural language processing (NLP) and computer vision (CV). The paradigm\npretrains a shared model on large volumes of unlabeled data and achieves\nstate-of-the-art (SOTA) for various tasks with minimal adaptation. However, the\nspeech processing community lacks a similar setup to systematically explore the\nparadigm. To bridge this gap, we introduce Speech processing Universal\nPERformance Benchmark (SUPERB). SUPERB is a leaderboard to benchmark the\nperformance of a shared model across a wide range of speech processing tasks\nwith minimal architecture changes and labeled data. Among multiple usages of the\nshared model, we especially focus on extracting the representation learned from\nSSL due to its preferable re-usability. We present a simple framework to solve\nSUPERB tasks by learning task-specialized lightweight prediction heads on top of\nthe frozen shared model. Our results demonstrate that the framework is promising\nas SSL representations show competitive generalizability and accessibility\nacross SUPERB tasks. We release SUPERB as a challenge with a leaderboard and a\nbenchmark toolkit to fuel the research in representation learning and general\nspeech processing.\n\nNote that in order to limit the required storage for preparing this dataset, the\naudio is stored in the .flac format and is not converted to a float32 array. To\nconvert, the audio file to a float32 array, please make use of the `.map()`\nfunction as follows:\n\n\n```python\nimport soundfile as sf\n\ndef map_to_array(batch):\n speech_array, _ = sf.read(batch[\"file\"])\n batch[\"speech\"] = speech_array\n return batch\n\ndataset = dataset.map(map_to_array, remove_columns=[\"file\"])\n```\n", "citation": "@article{DBLP:journals/corr/abs-2105-01051,\n author = {Shu{-}Wen Yang and\n Po{-}Han Chi and\n Yung{-}Sung Chuang and\n Cheng{-}I Jeff Lai and\n Kushal Lakhotia and\n Yist Y. Lin and\n Andy T. Liu and\n Jiatong Shi and\n Xuankai Chang and\n Guan{-}Ting Lin and\n Tzu{-}Hsien Huang and\n Wei{-}Cheng Tseng and\n Ko{-}tik Lee and\n Da{-}Rong Liu and\n Zili Huang and\n Shuyan Dong and\n Shang{-}Wen Li and\n Shinji Watanabe and\n Abdelrahman Mohamed and\n Hung{-}yi Lee},\n title = {{SUPERB:} Speech processing Universal PERformance Benchmark},\n journal = {CoRR},\n volume = {abs/2105.01051},\n year = {2021},\n url = {https://arxiv.org/abs/2105.01051},\n archivePrefix = {arXiv},\n eprint = {2105.01051},\n timestamp = {Thu, 01 Jul 2021 13:30:22 +0200},\n biburl = {https://dblp.org/rec/journals/corr/abs-2105-01051.bib},\n bibsource = {dblp computer science bibliography, https://dblp.org}\n}\n", "homepage": "https://www.tensorflow.org/datasets/catalog/speech_commands", "license": "", "features": {"file": {"dtype": "string", "id": null, "_type": "Value"}, "label": {"num_classes": 12, "names": ["yes", "no", "up", "down", "left", "right", "on", "off", "stop", "go", "_silence_", "_unknown_"], "names_file": null, "id": null, "_type": "ClassLabel"}}, "post_processed": null, "supervised_keys": {"input": "file", "output": "label"}, "task_templates": null, "builder_name": "superb", "config_name": "ks", "version": {"version_str": "1.9.0", "description": "", "major": 1, "minor": 9, "patch": 0}, "splits": {"train": {"name": "train", "num_bytes": 8467781, "num_examples": 51094, "dataset_name": "superb"}, "validation": {"name": "validation", "num_bytes": 1126476, "num_examples": 6798, "dataset_name": "superb"}, "test": {"name": "test", "num_bytes": 510619, "num_examples": 3081, "dataset_name": "superb"}}, "download_checksums": {"http://download.tensorflow.org/data/speech_commands_v0.01.tar.gz": {"num_bytes": 1489096277, "checksum": "743935421bb51cccdb6bdd152e04c5c70274e935c82119ad7faeec31780d811d"}, "http://download.tensorflow.org/data/speech_commands_test_set_v0.01.tar.gz": {"num_bytes": 71271436, "checksum": "baa084f6b62c91de660ff0588ae4dfc4e4d534aa99ac0e5f406cba75836cbd00"}}, "download_size": 1560367713, "post_processing_size": null, "dataset_size": 10104876, "size_in_bytes": 1570472589}}
dummy/ks/1.9.0/dummy_data.zip ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:eb2268a62de98190c456b9bc7eefdbd749b732a49068b54767e94250475aee08
3
+ size 4763
dummy/sd/1.9.0/dummy_data.zip ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6ba0c33b119a2c65fd6748fd623d14adde2ca70536fa5475f848316441b25c0a
3
+ size 3463
superb.py CHANGED
@@ -20,6 +20,7 @@
20
  import glob
21
  import os
22
  import textwrap
 
23
 
24
  import datasets
25
  from datasets.tasks import AutomaticSpeechRecognition
@@ -103,14 +104,18 @@ class SuperbConfig(datasets.BuilderConfig):
103
 
104
  def __init__(
105
  self,
 
106
  data_url,
107
  url,
 
108
  task_templates=None,
109
  **kwargs,
110
  ):
111
- super(SuperbConfig, self).__init__(version=datasets.Version("1.9.0", ""), **kwargs)
 
112
  self.data_url = data_url
113
  self.url = url
 
114
  self.task_templates = task_templates
115
 
116
 
@@ -129,15 +134,6 @@ class Superb(datasets.GeneratorBasedBuilder):
129
  training/validation/testing. The evaluation metric is word error
130
  rate (WER)."""
131
  ),
132
- url="http://www.openslr.org/12",
133
- data_url="http://www.openslr.org/resources/12/",
134
- task_templates=[AutomaticSpeechRecognition(audio_file_path_column="file", transcription_column="text")],
135
- )
136
- ]
137
-
138
- def _info(self):
139
- return datasets.DatasetInfo(
140
- description=_DESCRIPTION,
141
  features=datasets.Features(
142
  {
143
  "file": datasets.Value("string"),
@@ -148,6 +144,82 @@ class Superb(datasets.GeneratorBasedBuilder):
148
  }
149
  ),
150
  supervised_keys=("file", "text"),
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
151
  homepage=self.config.url,
152
  citation=_CITATION,
153
  task_templates=self.config.task_templates,
@@ -169,24 +241,235 @@ class Superb(datasets.GeneratorBasedBuilder):
169
  ),
170
  datasets.SplitGenerator(name=datasets.Split.TEST, gen_kwargs={"archive_path": archive_path["test"]}),
171
  ]
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
172
 
173
- def _generate_examples(self, archive_path):
174
  """Generate examples."""
175
- transcripts_glob = os.path.join(archive_path, "LibriSpeech", "*/*/*/*.txt")
176
- key = 0
177
- for transcript_path in sorted(glob.glob(transcripts_glob)):
178
- transcript_dir_path = os.path.dirname(transcript_path)
179
- with open(transcript_path, "r", encoding="utf-8") as f:
180
- for line in f:
181
- line = line.strip()
182
- id_, transcript = line.split(" ", 1)
183
- audio_file = f"{id_}.flac"
184
- speaker_id, chapter_id = [int(el) for el in id_.split("-")[:2]]
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
185
  yield key, {
186
- "id": id_,
187
- "speaker_id": speaker_id,
188
- "chapter_id": chapter_id,
189
- "file": os.path.join(transcript_dir_path, audio_file),
190
- "text": transcript,
191
  }
192
- key += 1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
20
  import glob
21
  import os
22
  import textwrap
23
+ from dataclasses import dataclass
24
 
25
  import datasets
26
  from datasets.tasks import AutomaticSpeechRecognition
 
104
 
105
  def __init__(
106
  self,
107
+ features,
108
  data_url,
109
  url,
110
+ supervised_keys=None,
111
  task_templates=None,
112
  **kwargs,
113
  ):
114
+ super().__init__(version=datasets.Version("1.9.0", ""), **kwargs)
115
+ self.features = features
116
  self.data_url = data_url
117
  self.url = url
118
+ self.supervised_keys = supervised_keys
119
  self.task_templates = task_templates
120
 
121
 
 
134
  training/validation/testing. The evaluation metric is word error
135
  rate (WER)."""
136
  ),
 
 
 
 
 
 
 
 
 
137
  features=datasets.Features(
138
  {
139
  "file": datasets.Value("string"),
 
144
  }
145
  ),
146
  supervised_keys=("file", "text"),
147
+ url="http://www.openslr.org/12",
148
+ data_url="http://www.openslr.org/resources/12/",
149
+ task_templates=[AutomaticSpeechRecognition(audio_file_path_column="file", transcription_column="text")],
150
+ ),
151
+ SuperbConfig(
152
+ name="ks",
153
+ description=textwrap.dedent(
154
+ """\
155
+ Keyword Spotting (KS) detects preregistered keywords by classifying utterances into a predefined set of
156
+ words. The task is usually performed on-device for the fast response time. Thus, accuracy, model size, and
157
+ inference time are all crucial. SUPERB uses the widely used [Speech Commands dataset v1.0] for the task.
158
+ The dataset consists of ten classes of keywords, a class for silence, and an unknown class to include the
159
+ false positive. The evaluation metric is accuracy (ACC)"""
160
+ ),
161
+ features=datasets.Features(
162
+ {
163
+ "file": datasets.Value("string"),
164
+ "label": datasets.ClassLabel(
165
+ names=[
166
+ "yes",
167
+ "no",
168
+ "up",
169
+ "down",
170
+ "left",
171
+ "right",
172
+ "on",
173
+ "off",
174
+ "stop",
175
+ "go",
176
+ "_silence_",
177
+ "_unknown_",
178
+ ]
179
+ ),
180
+ }
181
+ ),
182
+ supervised_keys=("file", "label"),
183
+ url="https://www.tensorflow.org/datasets/catalog/speech_commands",
184
+ data_url="http://download.tensorflow.org/data/{filename}",
185
+ ),
186
+ SuperbConfig(
187
+ name="sd",
188
+ description=textwrap.dedent(
189
+ """\
190
+ Speaker Diarization (SD) predicts `who is speaking when` for each timestamp, and multiple speakers can
191
+ speak simultaneously. The model has to encode rich speaker characteristics for each frame and should be
192
+ able to represent mixtures of signals. [LibriMix] is adopted where LibriSpeech
193
+ train-clean-100/dev-clean/test-clean are used to generate mixtures for training/validation/testing.
194
+ We focus on the two-speaker scenario as the first step. The time-coded speaker labels were generated using
195
+ alignments from Kaldi LibriSpeech ASR model. The evaluation metric is diarization error rate (DER)."""
196
+ ),
197
+ features=datasets.Features(
198
+ {
199
+ "record_id": datasets.Value("string"),
200
+ "file": datasets.Value("string"),
201
+ "start": datasets.Value("int64"),
202
+ "end": datasets.Value("int64"),
203
+ "speakers": [
204
+ {
205
+ "speaker_id": datasets.Value("string"),
206
+ "start": datasets.Value("int64"),
207
+ "end": datasets.Value("int64"),
208
+ }
209
+ ],
210
+ }
211
+ ), # TODO
212
+ supervised_keys=None, # TODO
213
+ url="https://github.com/ftshijt/LibriMix",
214
+ data_url="https://huggingface.co/datasets/superb/superb-data/resolve/main/sd/{split}/{filename}",
215
+ ),
216
+ ]
217
+
218
+ def _info(self):
219
+ return datasets.DatasetInfo(
220
+ description=_DESCRIPTION,
221
+ features=self.config.features,
222
+ supervised_keys=self.config.supervised_keys,
223
  homepage=self.config.url,
224
  citation=_CITATION,
225
  task_templates=self.config.task_templates,
 
241
  ),
242
  datasets.SplitGenerator(name=datasets.Split.TEST, gen_kwargs={"archive_path": archive_path["test"]}),
243
  ]
244
+ elif self.config.name == "ks":
245
+ _DL_URLS = {
246
+ "train_val_test": self.config.data_url.format(filename="speech_commands_v0.01.tar.gz"),
247
+ "test": self.config.data_url.format(filename="speech_commands_test_set_v0.01.tar.gz"),
248
+ }
249
+ archive_path = dl_manager.download_and_extract(_DL_URLS)
250
+ return [
251
+ datasets.SplitGenerator(
252
+ name=datasets.Split.TRAIN,
253
+ gen_kwargs={"archive_path": archive_path["train_val_test"], "split": "train"},
254
+ ),
255
+ datasets.SplitGenerator(
256
+ name=datasets.Split.VALIDATION,
257
+ gen_kwargs={"archive_path": archive_path["train_val_test"], "split": "val"},
258
+ ),
259
+ datasets.SplitGenerator(
260
+ name=datasets.Split.TEST, gen_kwargs={"archive_path": archive_path["test"], "split": "test"}
261
+ ),
262
+ ]
263
+ elif self.config.name == "sd":
264
+ splits = ["train", "dev", "test"]
265
+ _DL_URLS = {
266
+ split: {
267
+ filename: self.config.data_url.format(split=split, filename=filename)
268
+ for filename in ["reco2dur", "segments", "utt2spk", "wav.zip"]
269
+ }
270
+ for split in splits
271
+ }
272
+ archive_path = dl_manager.download_and_extract(_DL_URLS)
273
+ return [
274
+ datasets.SplitGenerator(
275
+ name=datasets.NamedSplit(split), gen_kwargs={"archive_path": archive_path[split], "split": split}
276
+ )
277
+ for split in splits
278
+ ]
279
 
280
+ def _generate_examples(self, archive_path, split=None):
281
  """Generate examples."""
282
+ if self.config.name == "asr":
283
+ transcripts_glob = os.path.join(archive_path, "LibriSpeech", "*/*/*/*.txt")
284
+ key = 0
285
+ for transcript_path in sorted(glob.glob(transcripts_glob)):
286
+ transcript_dir_path = os.path.dirname(transcript_path)
287
+ with open(transcript_path, "r", encoding="utf-8") as f:
288
+ for line in f:
289
+ line = line.strip()
290
+ id_, transcript = line.split(" ", 1)
291
+ audio_file = f"{id_}.flac"
292
+ speaker_id, chapter_id = [int(el) for el in id_.split("-")[:2]]
293
+ yield key, {
294
+ "id": id_,
295
+ "speaker_id": speaker_id,
296
+ "chapter_id": chapter_id,
297
+ "file": os.path.join(transcript_dir_path, audio_file),
298
+ "text": transcript,
299
+ }
300
+ key += 1
301
+ elif self.config.name == "ks":
302
+ words = ["yes", "no", "up", "down", "left", "right", "on", "off", "stop", "go"]
303
+ splits = _split_ks_files(archive_path, split)
304
+ for key, audio_file in enumerate(sorted(splits[split])):
305
+ base_dir, file_name = os.path.split(audio_file)
306
+ _, word = os.path.split(base_dir)
307
+ if word in words:
308
+ label = word
309
+ elif word == "_silence_" or word == "_background_noise_":
310
+ label = "_silence_"
311
+ else:
312
+ label = "_unknown_"
313
+ yield key, {"file": audio_file, "label": label}
314
+ elif self.config.name == "sd":
315
+ data = SdData(archive_path)
316
+ args = SdArgs()
317
+ chunk_indices = _generate_chunk_indices(data, args, split=split)
318
+ if split != "test":
319
+ for key, (rec, st, ed) in enumerate(chunk_indices):
320
+ speakers = _get_speakers(rec, data, args)
321
  yield key, {
322
+ "record_id": rec,
323
+ "file": data.wavs[rec],
324
+ "start": st,
325
+ "end": ed,
326
+ "speakers": speakers,
327
  }
328
+ else:
329
+ key = 0
330
+ for rec in chunk_indices:
331
+ for rec, st, ed in chunk_indices[rec]:
332
+ speakers = _get_speakers(rec, data, args)
333
+ yield key, {
334
+ "record_id": rec,
335
+ "file": data.wavs[rec],
336
+ "start": st,
337
+ "end": ed,
338
+ "speakers": speakers,
339
+ }
340
+ key += 1
341
+
342
+
343
+ class SdData:
344
+ def __init__(self, data_dir):
345
+ """Load sd data."""
346
+ self.segments = self._load_segments_rechash(data_dir["segments"])
347
+ self.utt2spk = self._load_utt2spk(data_dir["utt2spk"])
348
+ self.wavs = self._load_wav_zip(data_dir["wav.zip"])
349
+ self.reco2dur = self._load_reco2dur(data_dir["reco2dur"])
350
+
351
+ def _load_segments_rechash(self, segments_file):
352
+ """Load segments file as dict with recid index."""
353
+ ret = {}
354
+ if not os.path.exists(segments_file):
355
+ return None
356
+ with open(segments_file, encoding="utf-8") as f:
357
+ for line in f:
358
+ utt, rec, st, et = line.strip().split()
359
+ if rec not in ret:
360
+ ret[rec] = []
361
+ ret[rec].append({"utt": utt, "st": float(st), "et": float(et)})
362
+ return ret
363
+
364
+ def _load_wav_zip(self, wav_zip):
365
+ """Return dictionary { rec: wav_rxfilename }."""
366
+ wav_dir = os.path.join(wav_zip, "wav")
367
+ return {
368
+ os.path.splitext(filename)[0]: os.path.join(wav_dir, filename) for filename in sorted(os.listdir(wav_dir))
369
+ }
370
+
371
+ def _load_utt2spk(self, utt2spk_file):
372
+ """Returns dictionary { uttid: spkid }."""
373
+ with open(utt2spk_file, encoding="utf-8") as f:
374
+ lines = [line.strip().split(None, 1) for line in f]
375
+ return {x[0]: x[1] for x in lines}
376
+
377
+ def _load_reco2dur(self, reco2dur_file):
378
+ """Returns dictionary { recid: duration }."""
379
+ if not os.path.exists(reco2dur_file):
380
+ return None
381
+ with open(reco2dur_file, encoding="utf-8") as f:
382
+ lines = [line.strip().split(None, 1) for line in f]
383
+ return {x[0]: float(x[1]) for x in lines}
384
+
385
+
386
+ @dataclass
387
+ class SdArgs:
388
+ chunk_size: int = 2000
389
+ frame_shift: int = 160
390
+ subsampling: int = 1
391
+ label_delay: int = 0
392
+ num_speakers: int = 2
393
+ rate: int = 16000
394
+ use_last_samples: bool = True
395
+
396
+
397
+ def _generate_chunk_indices(data, args, split=None):
398
+ chunk_indices = [] if split != "test" else {}
399
+ # make chunk indices: filepath, start_frame, end_frame
400
+ for rec in data.wavs:
401
+ data_len = int(data.reco2dur[rec] * args.rate / args.frame_shift)
402
+ data_len = int(data_len / args.subsampling)
403
+ if split == "test":
404
+ chunk_indices[rec] = []
405
+ if split != "test":
406
+ for st, ed in _gen_frame_indices(
407
+ data_len,
408
+ args.chunk_size,
409
+ args.chunk_size,
410
+ args.use_last_samples,
411
+ label_delay=args.label_delay,
412
+ subsampling=args.subsampling,
413
+ ):
414
+ chunk_indices.append((rec, st * args.subsampling, ed * args.subsampling))
415
+ else:
416
+ for st, ed in _gen_chunk_indices(data_len, args.chunk_size):
417
+ chunk_indices[rec].append((rec, st * args.subsampling, ed * args.subsampling))
418
+ return chunk_indices
419
+
420
+
421
+ def _count_frames(data_len, size, step):
422
+ # no padding at edges, last remaining samples are ignored
423
+ return int((data_len - size + step) / step)
424
+
425
+
426
+ def _gen_frame_indices(data_length, size=2000, step=2000, use_last_samples=False, label_delay=0, subsampling=1):
427
+ i = -1
428
+ for i in range(_count_frames(data_length, size, step)):
429
+ yield i * step, i * step + size
430
+ if use_last_samples and i * step + size < data_length:
431
+ if data_length - (i + 1) * step - subsampling * label_delay > 0:
432
+ yield (i + 1) * step, data_length
433
+
434
+
435
+ def _gen_chunk_indices(data_len, chunk_size):
436
+ step = chunk_size
437
+ start = 0
438
+ while start < data_len:
439
+ end = min(data_len, start + chunk_size)
440
+ yield start, end
441
+ start += step
442
+
443
+
444
+ def _get_speakers(rec, data, args):
445
+ return [
446
+ {
447
+ "speaker_id": data.utt2spk[segment["utt"]],
448
+ "start": round(segment["st"] * args.rate / args.frame_shift),
449
+ "end": round(segment["et"] * args.rate / args.frame_shift),
450
+ }
451
+ for segment in data.segments[rec]
452
+ ]
453
+
454
+
455
+ def _split_ks_files(archive_path, split):
456
+ audio_path = os.path.join(archive_path, "**/*.wav")
457
+ audio_paths = glob.glob(audio_path)
458
+ if split == "test":
459
+ # use all available files for the test archive
460
+ return {"test": audio_paths}
461
+
462
+ val_list_file = os.path.join(archive_path, "validation_list.txt")
463
+ test_list_file = os.path.join(archive_path, "testing_list.txt")
464
+ with open(val_list_file, encoding="utf-8") as f:
465
+ val_paths = f.read().strip().splitlines()
466
+ val_paths = [os.path.join(archive_path, p) for p in val_paths]
467
+ with open(test_list_file, encoding="utf-8") as f:
468
+ test_paths = f.read().strip().splitlines()
469
+ test_paths = [os.path.join(archive_path, p) for p in test_paths]
470
+
471
+ # the paths for the train set is just whichever paths that do not exist in
472
+ # either the test or validation splits
473
+ train_paths = list(set(audio_paths) - set(val_paths) - set(test_paths))
474
+
475
+ return {"train": train_paths, "val": val_paths}