tobiolatunji commited on
Commit
75fceb6
1 Parent(s): 27520da

add configs for smaller datasets

Browse files
Files changed (2) hide show
  1. README.md +69 -0
  2. afrispeech-200.py +64 -7
README.md CHANGED
@@ -50,6 +50,7 @@ dataset_info:
50
  - [Table of Contents](#table-of-contents)
51
  - [Dataset Description](#dataset-description)
52
  - [Dataset Summary](#dataset-summary)
 
53
  - [Supported Tasks and Leaderboards](#supported-tasks-and-leaderboards)
54
  - [Languages](#languages)
55
  - [Dataset Structure](#dataset-structure)
@@ -88,6 +89,72 @@ dataset_info:
88
  AFRISPEECH-200 is a 200hr Pan-African speech corpus for clinical and general domain English accented ASR; a dataset with 120 African accents from 13 countries and 2,463 unique African speakers.
89
  Our goal is to raise awareness for and advance Pan-African English ASR research, especially for the clinical domain.
90
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
91
  ### Supported Tasks and Leaderboards
92
 
93
  - Automatic Speech Recognition
@@ -112,6 +179,8 @@ A typical data point comprises the path to the audio file, called `path` and its
112
  'transcirpt': 'The patient took the correct medication'}
113
  ```
114
 
 
 
115
  ### Data Fields
116
 
117
  - speaker_id: An id for which speaker (voice) made the recording
 
50
  - [Table of Contents](#table-of-contents)
51
  - [Dataset Description](#dataset-description)
52
  - [Dataset Summary](#dataset-summary)
53
+ - [How to use](#how-to-use)
54
  - [Supported Tasks and Leaderboards](#supported-tasks-and-leaderboards)
55
  - [Languages](#languages)
56
  - [Dataset Structure](#dataset-structure)
 
89
  AFRISPEECH-200 is a 200hr Pan-African speech corpus for clinical and general domain English accented ASR; a dataset with 120 African accents from 13 countries and 2,463 unique African speakers.
90
  Our goal is to raise awareness for and advance Pan-African English ASR research, especially for the clinical domain.
91
 
92
+ ## How to use
93
+
94
+ The `datasets` library allows you to load and pre-process your dataset in pure Python, at scale. The dataset can be downloaded and prepared in one call to your local drive by using the `load_dataset` function.
95
+
96
+ For example, to download the isizulu config, simply specify the corresponding language config name, list of supported accents provided in accent list section below:
97
+ ```python
98
+ from datasets import load_dataset
99
+
100
+ afrispeech = load_dataset("tobiolatunji/afrispeech-200", "isizulu", split="train")
101
+ ```
102
+
103
+ Using the datasets library, you can also stream the dataset on-the-fly by adding a `streaming=True` argument to the `load_dataset` function call. Loading a dataset in streaming mode loads individual samples of the dataset at a time, rather than downloading the entire dataset to disk.
104
+ ```python
105
+ from datasets import load_dataset
106
+
107
+ afrispeech = load_dataset("tobiolatunji/afrispeech-200", "isizulu", split="train", streaming=True)
108
+
109
+ print(next(iter(afrispeech)))
110
+ ```
111
+
112
+ ### Local
113
+
114
+ ```python
115
+ from datasets import load_dataset
116
+ from torch.utils.data.sampler import BatchSampler, RandomSampler
117
+
118
+ afrispeech = load_dataset("tobiolatunji/afrispeech-200", "isizulu", split="train")
119
+ batch_sampler = BatchSampler(RandomSampler(afrispeech), batch_size=32, drop_last=False)
120
+ dataloader = DataLoader(afrispeech, batch_sampler=batch_sampler)
121
+ ```
122
+
123
+ ### Streaming
124
+
125
+ ```python
126
+ from datasets import load_dataset
127
+ from torch.utils.data import DataLoader
128
+
129
+ afrispeech = load_dataset("tobiolatunji/afrispeech-200", "isizulu", split="train")
130
+ dataloader = DataLoader(afrispeech, batch_size=32)
131
+ ```
132
+
133
+ To find out more about loading and preparing audio datasets, head over to [hf.co/blog/audio-datasets](https://huggingface.co/blog/audio-datasets).
134
+
135
+ ### Example scripts
136
+
137
+ Train your own CTC or Seq2Seq Automatic Speech Recognition models on AfriSpeech-200 with `transformers` - [here](https://github.com/huggingface/transformers/tree/main/examples/pytorch/speech-recognition).
138
+
139
+ AfriSpeech-200 can be downloaded and used as follows:
140
+
141
+ ```py
142
+ from datasets import load_dataset
143
+
144
+ afrispeech = load_dataset("tobiolatunji/afrispeech-200", "isizulu") # for isizulu,
145
+ # to download all data for multi-accent fine-tuning uncomment following line
146
+ # afrispeech = load_dataset("tobiolatunji/afrispeech-200", "all")
147
+
148
+ # see structure
149
+ print(afrispeech)
150
+
151
+ # load audio sample on the fly
152
+ audio_input = afrispeech["train"][0]["audio"] # audio bytes
153
+ transcript = afrispeech["train"][0]["transcript"] # transcript
154
+
155
+ # use audio_input and text transcript to fine-tune your model for audio classification
156
+ ```
157
+
158
  ### Supported Tasks and Leaderboards
159
 
160
  - Automatic Speech Recognition
 
179
  'transcirpt': 'The patient took the correct medication'}
180
  ```
181
 
182
+
183
+
184
  ### Data Fields
185
 
186
  - speaker_id: An id for which speaker (voice) made the recording
afrispeech-200.py CHANGED
@@ -35,6 +35,30 @@ Our goal is to raise awareness for and advance Pan-African English ASR research,
35
  especially for the clinical domain.
36
  """
37
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
38
  _HOMEPAGE = "https://github.com/intron-innovation/AfriSpeech-Dataset-Paper"
39
 
40
  _LICENSE = "http://creativecommons.org/licenses/by-nc-sa/4.0/"
@@ -51,10 +75,35 @@ _SHARDS = {
51
  'test': 4
52
  }
53
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
54
 
55
  class AfriSpeech(datasets.GeneratorBasedBuilder):
56
  DEFAULT_WRITER_BATCH_SIZE = 1000
57
- VERSION = datasets.Version("1.1.0")
 
58
 
59
  def _info(self):
60
  description = _DESCRIPTION
@@ -93,7 +142,13 @@ class AfriSpeech(datasets.GeneratorBasedBuilder):
93
  # with the url replaced with path to local files.
94
  # By default the archives will be extracted and a path to a cached folder
95
  # where they are extracted is returned instead of the archive
96
-
 
 
 
 
 
 
97
  n_shards = _SHARDS
98
 
99
  audio_urls = {}
@@ -139,14 +194,16 @@ class AfriSpeech(datasets.GeneratorBasedBuilder):
139
  with open(meta_path, "r", encoding="utf-8") as f:
140
  reader = csv.DictReader(f)
141
  for row in tqdm(reader, desc="Reading metadata..."):
142
- row["speaker_id"] = row["user_ids"]
143
- audio_id = row["audio_paths"].split("/")[-1]
144
- # if data is incomplete, fill with empty values
145
- metadata[audio_id] = {field: row.get(field, "") for field in data_fields}
 
146
 
147
  for i, audio_archive in enumerate(archives):
148
  for filename, file in audio_archive:
149
- _, filename = os.path.split(filename)
 
150
  if filename in metadata:
151
  result = dict(metadata[filename])
152
  # set the audio feature and the path to the extracted file
 
35
  especially for the clinical domain.
36
  """
37
 
38
+ _ALL_CONFIGS = [
39
+ 'yoruba', 'igbo', 'swahili', 'ijaw', 'xhosa', 'twi', 'luhya',
40
+ 'igala', 'urhobo', 'hausa', 'kiswahili', 'zulu', 'isizulu',
41
+ 'venda and xitsonga', 'borana', 'afrikaans', 'setswana', 'idoma',
42
+ 'izon', 'chichewa', 'ebira', 'tshivenda', 'isixhosa',
43
+ 'kinyarwanda', 'tswana', 'luganda', 'luo', 'venda', 'dholuo',
44
+ 'akan (fante)', 'sepedi', 'kikuyu', 'isindebele',
45
+ 'luganda and kiswahili', 'akan', 'sotho', 'south african english',
46
+ 'sesotho', 'swahili ,luganda ,arabic', 'shona', 'damara',
47
+ 'southern sotho', 'luo, swahili', 'ateso', 'meru', 'siswati',
48
+ 'portuguese', 'esan', 'nasarawa eggon', 'ibibio', 'isoko',
49
+ 'pidgin', 'alago', 'nembe', 'ngas', 'kagoma', 'ikwere', 'fulani',
50
+ 'bette', 'efik', 'edo', 'hausa/fulani', 'bekwarra', 'epie',
51
+ 'afemai', 'benin', 'nupe', 'tiv', 'okrika', 'etsako', 'ogoni',
52
+ 'kubi', 'gbagyi', 'brass', 'oklo', 'ekene', 'ika', 'berom', 'jaba',
53
+ 'itsekiri', 'ukwuani', 'yala mbembe', 'afo', 'english', 'ebiobo',
54
+ 'igbo and yoruba', 'okirika', 'kalabari', 'ijaw(nembe)', 'anaang',
55
+ 'eggon', 'bini', 'yoruba, hausa', 'ekpeye', 'bajju', 'kanuri',
56
+ 'delta', 'khana', 'ogbia', 'mada', 'mwaghavul', 'angas', 'ikulu',
57
+ 'eleme', 'igarra', 'etche', 'agatu', 'bassa', 'jukun', 'urobo',
58
+ 'ibani', 'obolo', 'idah', 'eket', 'nyandang', 'estako', 'ishan',
59
+ 'bassa-nge/nupe', 'bagi', 'gerawa'
60
+ ]
61
+
62
  _HOMEPAGE = "https://github.com/intron-innovation/AfriSpeech-Dataset-Paper"
63
 
64
  _LICENSE = "http://creativecommons.org/licenses/by-nc-sa/4.0/"
 
75
  'test': 4
76
  }
77
 
78
+ class AfriSpeechConfig(datasets.BuilderConfig):
79
+ """BuilderConfig for afrispeech"""
80
+
81
+ def __init__(
82
+ self, name, description, homepage, data_url
83
+ ):
84
+ super(AfriSpeechConfig, self).__init__(
85
+ name=self.name,
86
+ version=datasets.Version("1.0.0", ""),
87
+ description=self.description,
88
+ )
89
+ self.name = name
90
+ self.description = description
91
+ self.homepage = homepage
92
+ self.data_url = data_url
93
+
94
+
95
+ def _build_config(name):
96
+ return AfriSpeechConfig(
97
+ name=name,
98
+ description=_DESCRIPTION,
99
+ homepage=_HOMEPAGE_URL,
100
+ data_url=_DATA_URL,
101
+ )
102
 
103
  class AfriSpeech(datasets.GeneratorBasedBuilder):
104
  DEFAULT_WRITER_BATCH_SIZE = 1000
105
+ VERSION = datasets.Version("1.0.0")
106
+ BUILDER_CONFIGS = [_build_config(name) for name in _ALL_CONFIGS + ["all"]]
107
 
108
  def _info(self):
109
  description = _DESCRIPTION
 
142
  # with the url replaced with path to local files.
143
  # By default the archives will be extracted and a path to a cached folder
144
  # where they are extracted is returned instead of the archive
145
+
146
+ langs = (
147
+ _ALL_CONFIGS
148
+ if self.config.name == "all"
149
+ else [self.config.name]
150
+ )
151
+
152
  n_shards = _SHARDS
153
 
154
  audio_urls = {}
 
194
  with open(meta_path, "r", encoding="utf-8") as f:
195
  reader = csv.DictReader(f)
196
  for row in tqdm(reader, desc="Reading metadata..."):
197
+ if (row['accent'] == self.config.name) or (self.config.name == 'all'):
198
+ row["speaker_id"] = row["user_ids"]
199
+ audio_id = "/".join(row["audio_paths"].split("/")[-2:])
200
+ # if data is incomplete, fill with empty values
201
+ metadata[audio_id] = {field: row.get(field, "") for field in data_fields}
202
 
203
  for i, audio_archive in enumerate(archives):
204
  for filename, file in audio_archive:
205
+ # _, filename = os.path.split(filename)
206
+ filename = "/".join(filename.split("/")[-2:])
207
  if filename in metadata:
208
  result = dict(metadata[filename])
209
  # set the audio feature and the path to the extracted file