osanseviero HF staff commited on
Commit
917610f
1 Parent(s): cf09a41

Create automatic_speech_recognition_dummy.py

Browse files
Files changed (1) hide show
  1. automatic_speech_recognition_dummy.py +167 -0
automatic_speech_recognition_dummy.py ADDED
@@ -0,0 +1,167 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # coding=utf-8
2
+ # Copyright 2021 The TensorFlow Datasets Authors and the HuggingFace Datasets Authors.
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
+ # Lint as: python3
16
+ """SUPERB: Speech processing Universal PERformance Benchmark."""
17
+ import glob
18
+ import os
19
+ import textwrap
20
+ import datasets
21
+ from datasets.tasks import AutomaticSpeechRecognition
22
+
23
+ _CITATION = """\
24
+ @article{DBLP:journals/corr/abs-2105-01051,
25
+ author = {Shu{-}Wen Yang and
26
+ Po{-}Han Chi and
27
+ Yung{-}Sung Chuang and
28
+ Cheng{-}I Jeff Lai and
29
+ Kushal Lakhotia and
30
+ Yist Y. Lin and
31
+ Andy T. Liu and
32
+ Jiatong Shi and
33
+ Xuankai Chang and
34
+ Guan{-}Ting Lin and
35
+ Tzu{-}Hsien Huang and
36
+ Wei{-}Cheng Tseng and
37
+ Ko{-}tik Lee and
38
+ Da{-}Rong Liu and
39
+ Zili Huang and
40
+ Shuyan Dong and
41
+ Shang{-}Wen Li and
42
+ Shinji Watanabe and
43
+ Abdelrahman Mohamed and
44
+ Hung{-}yi Lee},
45
+ title = {{SUPERB:} Speech processing Universal PERformance Benchmark},
46
+ journal = {CoRR},
47
+ volume = {abs/2105.01051},
48
+ year = {2021},
49
+ url = {https://arxiv.org/abs/2105.01051},
50
+ archivePrefix = {arXiv},
51
+ eprint = {2105.01051},
52
+ timestamp = {Thu, 01 Jul 2021 13:30:22 +0200},
53
+ biburl = {https://dblp.org/rec/journals/corr/abs-2105-01051.bib},
54
+ bibsource = {dblp computer science bibliography, https://dblp.org}
55
+ }
56
+ """
57
+
58
+ _DESCRIPTION = """\
59
+ Self-supervised learning (SSL) has proven vital for advancing research in
60
+ natural language processing (NLP) and computer vision (CV). The paradigm
61
+ pretrains a shared model on large volumes of unlabeled data and achieves
62
+ state-of-the-art (SOTA) for various tasks with minimal adaptation. However, the
63
+ speech processing community lacks a similar setup to systematically explore the
64
+ paradigm. To bridge this gap, we introduce Speech processing Universal
65
+ PERformance Benchmark (SUPERB). SUPERB is a leaderboard to benchmark the
66
+ performance of a shared model across a wide range of speech processing tasks
67
+ with minimal architecture changes and labeled data. Among multiple usages of the
68
+ shared model, we especially focus on extracting the representation learned from
69
+ SSL due to its preferable re-usability. We present a simple framework to solve
70
+ SUPERB tasks by learning task-specialized lightweight prediction heads on top of
71
+ the frozen shared model. Our results demonstrate that the framework is promising
72
+ as SSL representations show competitive generalizability and accessibility
73
+ across SUPERB tasks. We release SUPERB as a challenge with a leaderboard and a
74
+ benchmark toolkit to fuel the research in representation learning and general
75
+ speech processing.
76
+ Note that in order to limit the required storage for preparing this dataset, the
77
+ audio is stored in the .flac format and is not converted to a float32 array. To
78
+ convert, the audio file to a float32 array, please make use of the `.map()`
79
+ function as follows:
80
+ ```python
81
+ import soundfile as sf
82
+ def map_to_array(batch):
83
+ speech_array, _ = sf.read(batch["file"])
84
+ batch["speech"] = speech_array
85
+ return batch
86
+ dataset = dataset.map(map_to_array, remove_columns=["file"])
87
+ ```
88
+ """
89
+
90
+ class AsrDummybConfig(datasets.BuilderConfig):
91
+ """BuilderConfig for Superb."""
92
+ def __init__(
93
+ self,
94
+ data_url,
95
+ url,
96
+ task_templates=None,
97
+ **kwargs,
98
+ ):
99
+ super(AsrDummybConfig, self).__init__(
100
+ version=datasets.Version("1.9.0", ""), **kwargs
101
+ )
102
+ self.data_url = data_url
103
+ self.url = url
104
+ self.task_templates = task_templates
105
+
106
+ class AsrDummy(datasets.GeneratorBasedBuilder):
107
+ """Superb dataset."""
108
+ BUILDER_CONFIGS = [
109
+ AsrDummybConfig(
110
+ name="asr",
111
+ description=textwrap.dedent(
112
+ """\
113
+ ASR transcribes utterances into words. While PR analyzes the
114
+ improvement in modeling phonetics, ASR reflects the significance of
115
+ the improvement in a real-world scenario. LibriSpeech
116
+ train-clean-100/dev-clean/test-clean subsets are used for
117
+ training/validation/testing. The evaluation metric is word error
118
+ rate (WER)."""
119
+ ),
120
+ url="http://www.openslr.org/12",
121
+ data_url="http://www.openslr.org/resources/12/",
122
+ task_templates=[
123
+ AutomaticSpeechRecognition(
124
+ audio_file_path_column="file", transcription_column="text"
125
+ )
126
+ ],
127
+ )
128
+ ]
129
+
130
+ DEFAULT_CONFIG_NAME = "asr"
131
+ def _info(self):
132
+ return datasets.DatasetInfo(
133
+ description=_DESCRIPTION,
134
+ features=datasets.Features(
135
+ {
136
+ "id": datasets.Value("string"),
137
+ "file": datasets.Value("string"),
138
+ }
139
+ ),
140
+ supervised_keys=("file",),
141
+ homepage=self.config.url,
142
+ citation=_CITATION,
143
+ task_templates=self.config.task_templates,
144
+ )
145
+
146
+ def _split_generators(self, dl_manager):
147
+ DL_URLS = [
148
+ f"https://huggingface.co/datasets/Narsil/automatic_speech_recognition_dummy/raw/main/{i}.flac"
149
+ for i in range(1, 4)
150
+ ]
151
+ archive_path = dl_manager.download_and_extract(DL_URLS)
152
+ return [
153
+ datasets.SplitGenerator(
154
+ name=datasets.Split.TEST,
155
+ gen_kwargs={"archive_path": archive_path},
156
+ ),
157
+ ]
158
+
159
+ def _generate_examples(self, archive_path):
160
+ """Generate examples."""
161
+ for i, filename in enumerate(archive_path):
162
+ key = str(i)
163
+ example = {
164
+ "id": key,
165
+ "file": filename,
166
+ }
167
+ yield key, example