sanchit-gandhi
commited on
Commit
·
21b99e4
1
Parent(s):
765c78e
Make tedlium work in streaming mode
Browse files- tedlium.py +165 -59
tedlium.py
CHANGED
@@ -16,6 +16,8 @@
|
|
16 |
|
17 |
import os
|
18 |
import re
|
|
|
|
|
19 |
from pathlib import Path
|
20 |
|
21 |
import numpy as np
|
@@ -25,16 +27,18 @@ import datasets
|
|
25 |
from datasets.tasks import AutomaticSpeechRecognition
|
26 |
|
27 |
|
|
|
|
|
28 |
_LICENSE = "licensed under Creative Commons BY-NC-ND 3.0 (http://creativecommons.org/licenses/by-nc-nd/3.0/deed.en)"
|
29 |
|
30 |
|
31 |
class TedliumReleaseConfig(datasets.BuilderConfig):
|
32 |
"""BuilderConfig for a release of the TED-LIUM dataset."""
|
33 |
|
34 |
-
def __init__(self, *, url,
|
35 |
super(TedliumReleaseConfig, self).__init__(version=datasets.Version("1.0.1"), **kwargs)
|
36 |
self.url = url
|
37 |
-
self.
|
38 |
# List of split, path pairs containing the relative path within the
|
39 |
# extracted tarball to the data for each split.
|
40 |
self.split_paths = split_paths
|
@@ -63,11 +67,15 @@ def _make_builder_configs():
|
|
63 |
}
|
64 |
""",
|
65 |
url="https://www.openslr.org/7/",
|
66 |
-
|
|
|
|
|
|
|
|
|
67 |
split_paths=[
|
68 |
-
(datasets.Split.TRAIN,
|
69 |
-
(datasets.Split.VALIDATION,
|
70 |
-
(datasets.Split.TEST,
|
71 |
],
|
72 |
)
|
73 |
|
@@ -97,11 +105,15 @@ def _make_builder_configs():
|
|
97 |
}
|
98 |
""",
|
99 |
url="https://www.openslr.org/19/",
|
100 |
-
|
|
|
|
|
|
|
|
|
101 |
split_paths=[
|
102 |
-
(datasets.Split.TRAIN,
|
103 |
-
(datasets.Split.VALIDATION,
|
104 |
-
(datasets.Split.TEST,
|
105 |
],
|
106 |
)
|
107 |
|
@@ -109,7 +121,8 @@ def _make_builder_configs():
|
|
109 |
name="release3",
|
110 |
description="""\
|
111 |
This is the TED-LIUM corpus release 3, licensed under Creative Commons
|
112 |
-
BY-NC-ND 3.0.
|
|
|
113 |
|
114 |
All talks and text are property of TED Conferences LLC.
|
115 |
|
@@ -124,7 +137,7 @@ def _make_builder_configs():
|
|
124 |
- 452 hours of audio
|
125 |
- 2351 aligned automatic transcripts in STM format
|
126 |
- TEDLIUM 2 dev and test data: 19 TED talks in SPH format with
|
127 |
-
corresponding manual transcriptions
|
128 |
- Dictionary with pronunciations (159848 entries), same file as the one
|
129 |
included in TED-LIUM 2
|
130 |
- Selected monolingual data for language modeling from WMT12 publicly
|
@@ -132,11 +145,6 @@ def _make_builder_configs():
|
|
132 |
have been modified to get a tokenization more relevant for English
|
133 |
language
|
134 |
|
135 |
-
Two corpus distributions:
|
136 |
-
- the legacy one, on which the dev and test datasets are the same as in
|
137 |
-
TED-LIUM 2 (and TED-LIUM 1).
|
138 |
-
- the 'speaker adaptation' one, especially designed for experiments on
|
139 |
-
speaker adaptation.
|
140 |
""",
|
141 |
citation="""\
|
142 |
@inproceedings{hernandez2018tedlium3,
|
@@ -149,18 +157,54 @@ def _make_builder_configs():
|
|
149 |
}
|
150 |
""",
|
151 |
url="https://www.openslr.org/51/",
|
152 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
153 |
split_paths=[
|
154 |
-
(datasets.Split.
|
155 |
-
(datasets.Split.
|
156 |
-
|
157 |
-
# which are skipped by extraction (see above).
|
158 |
-
# Work around this by manually dereferencing the links here.
|
159 |
-
(datasets.Split.TRAIN, os.path.join("TEDLIUM_release-3", "data")),
|
160 |
],
|
161 |
)
|
162 |
|
163 |
-
return [release1, release2, release3]
|
164 |
|
165 |
|
166 |
class TedLium(datasets.GeneratorBasedBuilder):
|
@@ -192,46 +236,108 @@ class TedLium(datasets.GeneratorBasedBuilder):
|
|
192 |
)
|
193 |
|
194 |
def _split_generators(self, dl_manager):
|
195 |
-
|
|
|
|
|
196 |
splits = []
|
197 |
for split, path in self.config.split_paths:
|
198 |
-
kwargs = {
|
|
|
|
|
|
|
|
|
199 |
splits.append(datasets.SplitGenerator(name=split, gen_kwargs=kwargs))
|
200 |
return splits
|
201 |
|
202 |
-
def _generate_examples(self, filepath):
|
203 |
"""Generate examples from a TED-LIUM stm file."""
|
204 |
-
|
205 |
-
|
206 |
-
|
207 |
-
|
208 |
-
|
209 |
-
|
210 |
-
|
211 |
-
|
212 |
-
|
213 |
-
|
214 |
-
|
215 |
-
|
216 |
-
line
|
217 |
-
|
218 |
-
|
219 |
-
|
220 |
-
|
221 |
-
|
222 |
-
|
223 |
-
|
224 |
-
|
225 |
-
|
226 |
-
|
227 |
-
|
228 |
-
|
229 |
-
|
230 |
-
|
231 |
-
|
232 |
-
|
233 |
-
|
234 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
235 |
|
236 |
|
237 |
def _maybe_trim_suffix(transcript):
|
|
|
16 |
|
17 |
import os
|
18 |
import re
|
19 |
+
from collections import defaultdict
|
20 |
+
from io import BytesIO
|
21 |
from pathlib import Path
|
22 |
|
23 |
import numpy as np
|
|
|
27 |
from datasets.tasks import AutomaticSpeechRecognition
|
28 |
|
29 |
|
30 |
+
_DL_URL = "https://huggingface.co/datasets/LIUM/tedlium/resolve/main/"
|
31 |
+
|
32 |
_LICENSE = "licensed under Creative Commons BY-NC-ND 3.0 (http://creativecommons.org/licenses/by-nc-nd/3.0/deed.en)"
|
33 |
|
34 |
|
35 |
class TedliumReleaseConfig(datasets.BuilderConfig):
|
36 |
"""BuilderConfig for a release of the TED-LIUM dataset."""
|
37 |
|
38 |
+
def __init__(self, *, url, download_urls, split_paths, citation, **kwargs):
|
39 |
super(TedliumReleaseConfig, self).__init__(version=datasets.Version("1.0.1"), **kwargs)
|
40 |
self.url = url
|
41 |
+
self.download_urls = download_urls
|
42 |
# List of split, path pairs containing the relative path within the
|
43 |
# extracted tarball to the data for each split.
|
44 |
self.split_paths = split_paths
|
|
|
67 |
}
|
68 |
""",
|
69 |
url="https://www.openslr.org/7/",
|
70 |
+
download_urls={
|
71 |
+
"train": _DL_URL + os.path.join("TEDLIUM_release1", "train.tar.gz"),
|
72 |
+
"validation": _DL_URL + os.path.join("TEDLIUM_release1", "dev.tar.gz"),
|
73 |
+
"test": _DL_URL + os.path.join("TEDLIUM_release1", "test.tar.gz"),
|
74 |
+
},
|
75 |
split_paths=[
|
76 |
+
(datasets.Split.TRAIN, "train"),
|
77 |
+
(datasets.Split.VALIDATION, "dev"),
|
78 |
+
(datasets.Split.TEST, "test"),
|
79 |
],
|
80 |
)
|
81 |
|
|
|
105 |
}
|
106 |
""",
|
107 |
url="https://www.openslr.org/19/",
|
108 |
+
download_urls={
|
109 |
+
"train": _DL_URL + os.path.join("TEDLIUM_release2", "train.tar.gz"),
|
110 |
+
"validation": _DL_URL + os.path.join("TEDLIUM_release2", "dev.tar.gz"),
|
111 |
+
"test": _DL_URL + os.path.join("TEDLIUM_release2", "test.tar.gz"),
|
112 |
+
},
|
113 |
split_paths=[
|
114 |
+
(datasets.Split.TRAIN, "train"),
|
115 |
+
(datasets.Split.VALIDATION, "dev"),
|
116 |
+
(datasets.Split.TEST, "test"),
|
117 |
],
|
118 |
)
|
119 |
|
|
|
121 |
name="release3",
|
122 |
description="""\
|
123 |
This is the TED-LIUM corpus release 3, licensed under Creative Commons
|
124 |
+
BY-NC-ND 3.0. This is the 'legacy' version of the corpus, in which the dev and test datasets are the same as in
|
125 |
+
TED-LIUM 2 (and TED-LIUM 1).
|
126 |
|
127 |
All talks and text are property of TED Conferences LLC.
|
128 |
|
|
|
137 |
- 452 hours of audio
|
138 |
- 2351 aligned automatic transcripts in STM format
|
139 |
- TEDLIUM 2 dev and test data: 19 TED talks in SPH format with
|
140 |
+
corresponding manual transcriptions.
|
141 |
- Dictionary with pronunciations (159848 entries), same file as the one
|
142 |
included in TED-LIUM 2
|
143 |
- Selected monolingual data for language modeling from WMT12 publicly
|
|
|
145 |
have been modified to get a tokenization more relevant for English
|
146 |
language
|
147 |
|
|
|
|
|
|
|
|
|
|
|
148 |
""",
|
149 |
citation="""\
|
150 |
@inproceedings{hernandez2018tedlium3,
|
|
|
157 |
}
|
158 |
""",
|
159 |
url="https://www.openslr.org/51/",
|
160 |
+
download_urls={
|
161 |
+
"train": _DL_URL + os.path.join("TEDLIUM_release3", "legacy", "train.tar.gz"),
|
162 |
+
"validation": _DL_URL + os.path.join("TEDLIUM_release3", "legacy", "dev.tar.gz"),
|
163 |
+
"test": _DL_URL + os.path.join("TEDLIUM_release3", "legacy", "test.tar.gz"),
|
164 |
+
},
|
165 |
+
split_paths=[
|
166 |
+
(datasets.Split.TRAIN, "train"),
|
167 |
+
(datasets.Split.VALIDATION, "dev"),
|
168 |
+
(datasets.Split.TEST, "test"),
|
169 |
+
],
|
170 |
+
)
|
171 |
+
|
172 |
+
release3_speaker_adaptation = TedliumReleaseConfig(
|
173 |
+
name="release3-speaker-adaptation",
|
174 |
+
description="""\
|
175 |
+
This is the TED-LIUM corpus release 3, licensed under Creative Commons
|
176 |
+
BY-NC-ND 3.0. This is the 'speaker adaptation' version of the corpus, specially designed for experiments on
|
177 |
+
speaker adaptation.
|
178 |
+
|
179 |
+
All talks and text are property of TED Conferences LLC.
|
180 |
+
|
181 |
+
This new TED-LIUM release was made through a collaboration between the
|
182 |
+
Ubiqus company and the LIUM (University of Le Mans, France)
|
183 |
+
""",
|
184 |
+
citation="""\
|
185 |
+
@inproceedings{hernandez2018tedlium3,
|
186 |
+
title={TED-LIUM 3: twice as much data and corpus repartition for experiments on speaker adaptation},
|
187 |
+
author={Hernandez, Fran{\\c{c}}ois and Nguyen, Vincent and Ghannay, Sahar and Tomashenko, Natalia and Est{\\`e}ve, Yannick},
|
188 |
+
booktitle={International Conference on Speech and Computer},
|
189 |
+
pages={198--208},
|
190 |
+
year={2018},
|
191 |
+
organization={Springer}
|
192 |
+
}
|
193 |
+
""",
|
194 |
+
url="https://www.openslr.org/51/",
|
195 |
+
download_urls={
|
196 |
+
"train": _DL_URL + os.path.join("TEDLIUM_release3", "speaker-adaptation", "train.tar.gz"),
|
197 |
+
"validation": _DL_URL + os.path.join("TEDLIUM_release3", "speaker-adaptation", "dev.tar.gz"),
|
198 |
+
"test": _DL_URL + os.path.join("TEDLIUM_release3", "speaker-adaptation", "test.tar.gz"),
|
199 |
+
},
|
200 |
split_paths=[
|
201 |
+
(datasets.Split.TRAIN, "train"),
|
202 |
+
(datasets.Split.VALIDATION, "dev"),
|
203 |
+
(datasets.Split.TEST, "test"),
|
|
|
|
|
|
|
204 |
],
|
205 |
)
|
206 |
|
207 |
+
return [release1, release2, release3, release3_speaker_adaptation]
|
208 |
|
209 |
|
210 |
class TedLium(datasets.GeneratorBasedBuilder):
|
|
|
236 |
)
|
237 |
|
238 |
def _split_generators(self, dl_manager):
|
239 |
+
archive_path = dl_manager.download(self.config.download_urls)
|
240 |
+
# (Optional) In non-streaming mode, we can extract the archive locally to have actual local audio files:
|
241 |
+
local_extracted_archive = dl_manager.extract(archive_path) if not dl_manager.is_streaming else {}
|
242 |
splits = []
|
243 |
for split, path in self.config.split_paths:
|
244 |
+
kwargs = {
|
245 |
+
"filepath": dl_manager.iter_archive(archive_path[split]),
|
246 |
+
"local_extracted_archive": local_extracted_archive.get(split),
|
247 |
+
"split_path": path,
|
248 |
+
}
|
249 |
splits.append(datasets.SplitGenerator(name=split, gen_kwargs=kwargs))
|
250 |
return splits
|
251 |
|
252 |
+
def _generate_examples(self, filepath, local_extracted_archive, split_path):
|
253 |
"""Generate examples from a TED-LIUM stm file."""
|
254 |
+
if local_extracted_archive:
|
255 |
+
# The stm directory houses the speaker and transcription information in .stm format
|
256 |
+
stm_dir = os.path.join(local_extracted_archive, split_path, "stm")
|
257 |
+
# The sph directory houses the audio files in .sph format
|
258 |
+
sph_dir = os.path.join(local_extracted_archive, split_path, "sph")
|
259 |
+
stm_files = [os.path.join(stm_dir, f) for f in os.listdir(stm_dir) if f.endswith(".stm")]
|
260 |
+
for file in stm_files:
|
261 |
+
# the .sph speaker file almost always has the same file name as the .stm file
|
262 |
+
speaker_file = Path(file).stem
|
263 |
+
audio_file = os.path.join(sph_dir, speaker_file + ".sph")
|
264 |
+
segment, sampling_rate = sf.read(audio_file, dtype=np.int16)
|
265 |
+
with open(file) as f:
|
266 |
+
for line in f:
|
267 |
+
line = line.strip()
|
268 |
+
fn, channel, speaker, start, end, label, transcript = line.split(" ", 6)
|
269 |
+
transcript = _maybe_trim_suffix(transcript)
|
270 |
+
if speaker_file != fn:
|
271 |
+
# handle the case where the stm file does not have the same file name as the transcript
|
272 |
+
speaker_file = fn
|
273 |
+
audio_file = os.path.join(sph_dir, speaker_file + ".sph")
|
274 |
+
segment, sampling_rate = sf.read(audio_file, dtype=np.int16)
|
275 |
+
samples = _extract_audio_segment(segment, int(channel), float(start), float(end))
|
276 |
+
key = "-".join([speaker, start, end, label])
|
277 |
+
example = {
|
278 |
+
"audio": {"path": audio_file, "array": samples, "sampling_rate": sampling_rate},
|
279 |
+
"text": transcript,
|
280 |
+
"speaker_id": speaker,
|
281 |
+
"gender": _parse_gender(label),
|
282 |
+
"file": audio_file,
|
283 |
+
"id": key,
|
284 |
+
}
|
285 |
+
yield key, example
|
286 |
+
|
287 |
+
else:
|
288 |
+
audio_data = {}
|
289 |
+
transcripts = defaultdict(list)
|
290 |
+
for path, f in filepath:
|
291 |
+
if path.endswith(".sph"):
|
292 |
+
# get the speaker id
|
293 |
+
fn = path.split("/")[-1].strip(".sph")
|
294 |
+
# read the audio data from raw byte form and add key-value pair to dict
|
295 |
+
audio_data[fn] = sf.read(BytesIO(f.read()), dtype=np.int16)
|
296 |
+
elif path.endswith(".stm"):
|
297 |
+
for line in f:
|
298 |
+
if line:
|
299 |
+
line = line.decode("utf-8").strip()
|
300 |
+
fn, channel, speaker, start, end, label, transcript = line.split(" ", 6)
|
301 |
+
transcript = _maybe_trim_suffix(transcript)
|
302 |
+
audio_file = path.replace("stm", "sph")
|
303 |
+
key = "-".join([speaker, start, end, label])
|
304 |
+
# append metadata information to the dict of transcripts for the associated speaker
|
305 |
+
transcripts[fn].append(
|
306 |
+
{
|
307 |
+
"text": transcript,
|
308 |
+
"speaker_id": speaker,
|
309 |
+
"gender": _parse_gender(label),
|
310 |
+
"file": audio_file,
|
311 |
+
"id": key,
|
312 |
+
"start": start,
|
313 |
+
"end": end,
|
314 |
+
"channel": channel,
|
315 |
+
"fn": fn,
|
316 |
+
}
|
317 |
+
)
|
318 |
+
|
319 |
+
if audio_data and audio_data.keys() == transcripts.keys():
|
320 |
+
for fn, speaker in transcripts.items():
|
321 |
+
for transcript in speaker:
|
322 |
+
segment, sampling_rate = audio_data[transcript["fn"]]
|
323 |
+
samples = _extract_audio_segment(
|
324 |
+
segment,
|
325 |
+
int(transcript["channel"]),
|
326 |
+
float(transcript["start"]),
|
327 |
+
float(transcript["end"]),
|
328 |
+
)
|
329 |
+
audio = {"path": transcript["file"], "array": samples, "sampling_rate": sampling_rate}
|
330 |
+
key = transcript["id"]
|
331 |
+
yield key, {
|
332 |
+
"audio": audio,
|
333 |
+
"text": transcript["text"],
|
334 |
+
"speaker_id": transcript["speaker_id"],
|
335 |
+
"gender": transcript["gender"],
|
336 |
+
"file": transcript["file"],
|
337 |
+
"id": transcript["id"],
|
338 |
+
}
|
339 |
+
audio_data = {}
|
340 |
+
transcripts = defaultdict(list)
|
341 |
|
342 |
|
343 |
def _maybe_trim_suffix(transcript):
|