holylovenia commited on
Commit
cac148f
1 Parent(s): 4cc27a8

Upload national_speech_corpus_sg_imda.py with huggingface_hub

Browse files
Files changed (1) hide show
  1. national_speech_corpus_sg_imda.py +301 -0
national_speech_corpus_sg_imda.py ADDED
@@ -0,0 +1,301 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os
2
+ import zipfile
3
+ from pathlib import Path
4
+ from typing import Dict, List, Tuple
5
+
6
+ try:
7
+ import audiosegment
8
+ except:
9
+ print("Please install audiosegment to use the `national_speech_corpus_sg_imda` dataloader.")
10
+ import datasets
11
+ import pandas as pd
12
+
13
+ try:
14
+ import textgrid
15
+ except:
16
+ print("Please install textgrid to use the `national_speech_corpus_sg_imda` dataloader.")
17
+
18
+ from seacrowd.utils import schemas
19
+ from seacrowd.utils.configs import SEACrowdConfig
20
+ from seacrowd.utils.constants import Licenses, Tasks
21
+
22
+ _CITATION = """\
23
+ @inproceedings{koh19_interspeech,
24
+ author={Jia Xin Koh and Aqilah Mislan and Kevin Khoo and Brian Ang and Wilson Ang and Charmaine Ng and Ying-Ying Tan},
25
+ title={{Building the Singapore English National Speech Corpus}},
26
+ year=2019,
27
+ booktitle={Proc. Interspeech 2019},
28
+ pages={321--325},
29
+ doi={10.21437/Interspeech.2019-1525},
30
+ issn={2308-457X}
31
+ }
32
+ """
33
+
34
+ _DATASETNAME = "national_speech_corpus_sg_imda"
35
+
36
+ _DESCRIPTION = """\
37
+ The National Speech Corpus (NSC) is the first large-scale Singapore English corpus spearheaded by the Info-communications and Media Development Authority (IMDA) of Singapore.
38
+ It aims to become an important source of open speech data for automatic speech recognition (ASR) research and speech-related applications.
39
+ The NSC improves speech engines’ accuracy of recognition and transcription for locally accented English.
40
+ The NSC is also able to contribute to speech synthesis technology where an AI voice can be produced that is more familiar to Singaporeans, with local terms pronounced more accurately.
41
+ """
42
+
43
+ _HOMEPAGE = "https://www.imda.gov.sg/how-we-can-help/national-speech-corpus"
44
+
45
+ _LANGUAGES = ["eng"] # We follow ISO639-3 language code (https://iso639-3.sil.org/code_tables/639/data)
46
+
47
+ _LICENSE = f"{Licenses.OTHERS.value} | Singapore Open Data Licence V1.0"
48
+
49
+ _LOCAL = True
50
+
51
+ _URLS = {}
52
+
53
+ # paths of all file locations, presented in a list to support different operating systems
54
+ _PATHS = {
55
+ "read_balanced": {
56
+ "metadata": ["PART1", "DOC", "Speaker Information (Part 1).XLSX"],
57
+ "data": {
58
+ "standing_mic": {
59
+ "audio": ["PART1", "DATA", "CHANNEL0", "WAVE"],
60
+ "text": ["PART1", "DATA", "CHANNEL0", "SCRIPT"],
61
+ },
62
+ "boundary_mic": {
63
+ "audio": ["PART1", "DATA", "CHANNEL1", "WAVE"],
64
+ "text": ["PART1", "DATA", "CHANNEL1", "SCRIPT"],
65
+ },
66
+ "phone": {
67
+ "audio": ["PART1", "DATA", "CHANNEL2", "WAVE"],
68
+ "text": ["PART1", "DATA", "CHANNEL2", "SCRIPT"],
69
+ },
70
+ },
71
+ },
72
+ "read_pertinent": {
73
+ "metadata": ["PART2", "DOC", "Speaker Information (Part 2).XLSX"],
74
+ "data": {
75
+ "standing_mic": {
76
+ "audio": ["PART2", "DATA", "CHANNEL0", "WAVE"],
77
+ "text": ["PART2", "DATA", "CHANNEL0", "SCRIPT"],
78
+ },
79
+ "boundary_mic": {
80
+ "audio": ["PART2", "DATA", "CHANNEL1", "WAVE"],
81
+ "text": ["PART2", "DATA", "CHANNEL1", "SCRIPT"],
82
+ },
83
+ "phone": {
84
+ "audio": ["PART2", "DATA", "CHANNEL2", "WAVE"],
85
+ "text": ["PART2", "DATA", "CHANNEL2", "SCRIPT"],
86
+ },
87
+ },
88
+ },
89
+ "conversational_f2f": {
90
+ "metadata": ["PART3", "Documents", "Speakers (Part 3).XLSX"],
91
+ "data": {
92
+ "close_mic": {
93
+ "audio": ["PART3", "Audio Same CloseMic"],
94
+ "text": ["PART3", "Scripts Same"],
95
+ },
96
+ "boundary_mic": {
97
+ "audio": ["PART3", "Audio Same BoundaryMic"],
98
+ "text": ["PART3", "Scripts Same"],
99
+ },
100
+ },
101
+ },
102
+ "conversational_telephone": {
103
+ "metadata": ["PART3", "Documents", "Speakers (Part 3).XLSX"],
104
+ "data": {
105
+ "ivr": {
106
+ "audio": ["PART3", "Audio Separate IVR"],
107
+ "text": ["PART3", "Scripts Separate"],
108
+ },
109
+ "standing_mic": {
110
+ "audio": ["PART3", "Audio Separate StandingMic"],
111
+ "text": ["PART3", "Scripts Separate"],
112
+ },
113
+ },
114
+ },
115
+ }
116
+
117
+ _SUPPORTED_TASKS = [Tasks.SPEECH_RECOGNITION]
118
+
119
+ _SOURCE_VERSION = "2.0.8" # should be 2.08 but HuggingFace does not allow
120
+
121
+ _SEACROWD_VERSION = "2024.06.20"
122
+
123
+
124
+ class NationalSpeechCorpusSgIMDADataset(datasets.GeneratorBasedBuilder):
125
+ """The National Speech Corpus (NSC), the first large-scale Singapore English corpus spearheaded by the Info-communications and Media Development Authority (IMDA) of Singapore."""
126
+
127
+ SOURCE_VERSION = datasets.Version(_SOURCE_VERSION)
128
+ SEACROWD_VERSION = datasets.Version(_SEACROWD_VERSION)
129
+
130
+ BUILDER_CONFIGS = [
131
+ SEACrowdConfig(
132
+ name="national_speech_corpus_sg_imda_source",
133
+ version=SOURCE_VERSION,
134
+ description="national_speech_corpus_sg_imda source schema",
135
+ schema="source",
136
+ subset_id="national_speech_corpus_sg_imda",
137
+ ),
138
+ SEACrowdConfig(
139
+ name="national_speech_corpus_sg_imda_seacrowd_sptext",
140
+ version=SEACROWD_VERSION,
141
+ description="national_speech_corpus_sg_imda SEACrowd schema",
142
+ schema="seacrowd_sptext",
143
+ subset_id="national_speech_corpus_sg_imda",
144
+ ),
145
+ ]
146
+
147
+ DEFAULT_CONFIG_NAME = "national_speech_corpus_sg_imda_source"
148
+
149
+ def _info(self) -> datasets.DatasetInfo:
150
+ if self.config.schema == "source":
151
+ features = datasets.Features(
152
+ {
153
+ "id": datasets.Value("string"),
154
+ "speaker_id": datasets.Value("string"),
155
+ "path": datasets.Value("string"),
156
+ "audio": datasets.Audio(sampling_rate=16_000),
157
+ "text": datasets.Value("string"),
158
+ }
159
+ )
160
+ elif self.config.schema == "seacrowd_sptext":
161
+ features = schemas.speech_text_features
162
+
163
+ return datasets.DatasetInfo(
164
+ description=_DESCRIPTION,
165
+ features=features,
166
+ homepage=_HOMEPAGE,
167
+ license=_LICENSE,
168
+ citation=_CITATION,
169
+ )
170
+
171
+ def _split_generators(self, dl_manager: datasets.DownloadManager) -> List[datasets.SplitGenerator]:
172
+ """Returns SplitGenerators."""
173
+ if self.config.data_dir is None:
174
+ raise ValueError(f"This is a local dataset. Please download the data from {_HOMEPAGE} and pass the path as the data_dir kwarg to load_dataset.")
175
+ else:
176
+ data_dir = self.config.data_dir
177
+
178
+ splits = []
179
+ for split_name, data in _PATHS.items():
180
+ metadata_path = os.path.join(*data["metadata"])
181
+ for mic_type, data_path in data["data"].items():
182
+ splits.append(
183
+ datasets.SplitGenerator(
184
+ name=f"{split_name}_{mic_type}",
185
+ gen_kwargs={"filepath": data_dir, "audio_dir": os.path.join(*data_path["audio"]), "text_dir": os.path.join(*data_path["text"]), "split": split_name, "mic_type": mic_type, "metadata_path": metadata_path},
186
+ )
187
+ )
188
+ return splits
189
+
190
+ def _read_part1_part2_text(self, text_path):
191
+ text_data = {}
192
+ with open(text_path, encoding="utf-8-sig") as f:
193
+ for line in f:
194
+ comp = line.split("\t")
195
+ text_id = comp[0].strip()
196
+ if len(text_id) > 1:
197
+ text_data[text_id] = comp[1].strip()
198
+ return text_data
199
+
200
+ def _generate_examples(self, filepath: Path, audio_dir: Path, text_dir: Path, split: str, mic_type: str, metadata_path: str, tmp_cache="~/.cache/nsc") -> Tuple[int, Dict]:
201
+ """Yields examples as (key, example) tuples."""
202
+
203
+ # get speaker info from Excel file
204
+ all_speaker_info = {}
205
+ if split.startswith("read"):
206
+ excel_file = pd.read_excel(open(os.path.join(filepath, metadata_path), "rb"), sheet_name="Speakers", dtype="object")
207
+ column_name = "SCD/PART" + "1" if split == "read_balanced" else 2
208
+ for idx, row in excel_file.iterrows():
209
+ all_speaker_info[str(row[column_name])] = {"gender": row["SEX"]}
210
+ elif split == "conversational_f2f":
211
+ excel_file = pd.read_excel(open(os.path.join(filepath, metadata_path), "rb"), sheet_name="Same Room", dtype="object")
212
+ for idx, row in excel_file.iterrows():
213
+ all_speaker_info[row["SCD"]] = {"age": row["AGE"], "gender": row["SEX"]}
214
+ elif split == "conversational_telephone":
215
+ excel_file = pd.read_excel(open(os.path.join(filepath, metadata_path), "rb"), sheet_name="Separate Room", dtype="object")
216
+ for idx, row in excel_file.iterrows():
217
+ all_speaker_info[f"{row['Conference ID']}_{row['Speaker ID']}"] = {"age": row["Age"], "gender": row["Gender"]}
218
+
219
+ for rel_text_path in os.listdir(os.path.join(filepath, text_dir)):
220
+ text_path = os.path.join(filepath, text_dir, rel_text_path)
221
+ text_name, _ext = os.path.splitext(rel_text_path)
222
+ if not os.path.isfile(text_path):
223
+ continue
224
+ if split.startswith("conversational"):
225
+ if split == "conversational_telephone":
226
+ # conf_{confid}_{confid}_{speaker_id} -> {confid}_{speaker_id}
227
+ speaker_id = text_name.split("_", 2)[-1]
228
+ pass
229
+ else:
230
+ speaker_id = text_name
231
+ speaker_info = all_speaker_info.get(speaker_id, {})
232
+ speaker_info["id"] = speaker_id
233
+
234
+ if mic_type in ["close_mic", "standing_mic"]:
235
+ audio_filename = text_name
236
+ audio_filepath = audio_filename + ".wav"
237
+ audio_path = os.path.join(audio_dir, audio_filepath)
238
+ elif mic_type == "boundary_mic":
239
+ audio_filename = text_name.split("-")[0]
240
+ audio_filepath = audio_filename + ".wav"
241
+ audio_path = os.path.join(audio_dir, audio_filepath)
242
+ elif mic_type == "ivr":
243
+ audio_subdir, audio_filename = text_name.rsplit("_", 1)
244
+ audio_filepath = audio_filename + ".wav"
245
+ audio_path = os.path.join(audio_dir, audio_subdir, audio_filepath)
246
+
247
+ audio_file = audiosegment.from_file(os.path.join(filepath, audio_path)).resample(sample_rate_Hz=16000)
248
+ text_spans = textgrid.TextGrid.fromFile(text_path)
249
+ audio_name, _ = os.path.splitext(os.path.basename(audio_dir))
250
+ for text_span in text_spans[0]:
251
+ start, end, text = text_span.minTime, text_span.maxTime, text_span.mark
252
+ key = f"{audio_name}_{start}_{end}"
253
+
254
+ start_sec, end_sec = int(start * 1000), int(end * 1000)
255
+ segment = audio_file[start_sec:end_sec]
256
+ export_dir = os.path.join(tmp_cache, "segmented", audio_name)
257
+ os.makedirs(export_dir, exist_ok=True)
258
+ segement_filename = os.path.join(export_dir, f"{audio_filename}-{round(start, 0)}-{round(end, 0)}.wav")
259
+ segment.export(segement_filename, format="wav")
260
+
261
+ example = {
262
+ "id": key,
263
+ "speaker_id": speaker_info["id"],
264
+ "path": segement_filename,
265
+ "audio": segement_filename,
266
+ "text": text,
267
+ }
268
+ if self.config.schema == "seacrowd_sptext":
269
+ example["metadata"] = {
270
+ "speaker_gender": speaker_info.get("gender", None), # not all speaker details are available
271
+ "speaker_age": None, # speaker age only available in age groups, but SEACrowd schema requires int64
272
+ }
273
+ yield key, example
274
+
275
+ else:
276
+ text_data = self._read_part1_part2_text(text_path)
277
+ audio_name, session = text_name[1:-1], text_name[-1]
278
+ speaker_id = audio_name
279
+ speaker_info = all_speaker_info.get(speaker_id, {})
280
+ speaker_info["id"] = speaker_id
281
+ for text_id, text in text_data.items():
282
+ with zipfile.ZipFile(os.path.join(filepath, audio_dir, f"SPEAKER{audio_name}.zip")) as zip_file:
283
+ zip_path = os.path.join(f"SPEAKER{audio_name}", f"SESSION{session}", f"{text_id}.WAV")
284
+ extract_path = os.path.join(tmp_cache, audio_dir)
285
+ os.makedirs(extract_path, exist_ok=True)
286
+ audio_path = zip_file.extract(zip_path, path=extract_path)
287
+
288
+ key = text_id
289
+ example = {
290
+ "id": key,
291
+ "speaker_id": speaker_info["id"],
292
+ "path": audio_path,
293
+ "audio": audio_path,
294
+ "text": text,
295
+ }
296
+ if self.config.schema == "seacrowd_sptext":
297
+ example["metadata"] = {
298
+ "speaker_gender": speaker_info.get("gender", None),
299
+ "speaker_age": None, # speaker age only available in age groups, but SEACrowd schema requires int64
300
+ }
301
+ yield key, example