Datasets:
Tasks:
Automatic Speech Recognition
Languages:
Swedish
update mix handling code
Browse files- waxholm.py +566 -51
waxholm.py
CHANGED
@@ -1,6 +1,6 @@
|
|
1 |
# coding=utf-8
|
2 |
# Copyright 2021 The TensorFlow Datasets Authors and the HuggingFace Datasets Authors.
|
3 |
-
# Copyright 2022 Jim O'Regan for Språkbanken Tal
|
4 |
#
|
5 |
# Licensed under the Apache License, Version 2.0 (the "License");
|
6 |
# you may not use this file except in compliance with the License.
|
@@ -20,6 +20,9 @@
|
|
20 |
from io import BytesIO
|
21 |
import os
|
22 |
import soundfile as sf
|
|
|
|
|
|
|
23 |
|
24 |
import datasets
|
25 |
from datasets.tasks import AutomaticSpeechRecognition
|
@@ -54,6 +57,13 @@ _CITATION = """
|
|
54 |
_URL = "http://www.speech.kth.se/waxholm/waxholm2.html"
|
55 |
|
56 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
57 |
class WaxholmDataset(datasets.GeneratorBasedBuilder):
|
58 |
"""Dataset script for Waxholm."""
|
59 |
|
@@ -68,6 +78,7 @@ class WaxholmDataset(datasets.GeneratorBasedBuilder):
|
|
68 |
{
|
69 |
"id": datasets.Value("string"),
|
70 |
"text": datasets.Value("string"),
|
|
|
71 |
"audio": datasets.Audio(sampling_rate=16_000)
|
72 |
}
|
73 |
)
|
@@ -79,7 +90,7 @@ class WaxholmDataset(datasets.GeneratorBasedBuilder):
|
|
79 |
homepage=_URL,
|
80 |
citation=_CITATION,
|
81 |
task_templates=[
|
82 |
-
AutomaticSpeechRecognition(
|
83 |
],
|
84 |
)
|
85 |
|
@@ -120,46 +131,81 @@ class WaxholmDataset(datasets.GeneratorBasedBuilder):
|
|
120 |
buffer = BytesIO()
|
121 |
sf.write(buffer, samples, sr, format="wav")
|
122 |
blank = Audio()
|
123 |
-
audio_to_pass = blank.encode_example(value = {"bytes": buffer.getvalue(), "sampling_rate": sr, })
|
124 |
yield line, {
|
125 |
"id": line,
|
126 |
"text": mix.text,
|
127 |
-
"
|
|
|
|
|
|
|
|
|
128 |
}
|
129 |
|
130 |
|
131 |
def fix_text(text: str) -> str:
|
132 |
-
replacements = text.maketrans("{}|\\", "
|
133 |
return text.translate(replacements)
|
134 |
|
135 |
|
|
|
|
|
|
|
136 |
class FR:
|
137 |
-
def __init__(self, text
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
138 |
if not text.startswith("FR"):
|
139 |
-
raise
|
140 |
parts = [a.strip() for a in text.split("\t")]
|
141 |
self.frame = parts[0][2:].strip()
|
142 |
if parts[-1].strip().endswith(" sec"):
|
143 |
self.seconds = parts[-1].strip()[0:-4]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
144 |
for subpart in parts[1:-1]:
|
145 |
-
|
146 |
-
|
147 |
-
|
148 |
-
|
149 |
-
|
150 |
-
|
151 |
-
|
152 |
-
self.phone = fix_text(subpart[2:])
|
153 |
-
elif subpart.startswith("#"):
|
154 |
-
self.type = 'B'
|
155 |
-
self.phone_type = fix_text(subpart[0:2])
|
156 |
-
self.phone = fix_text(subpart[2:])
|
157 |
elif subpart.startswith(">pm "):
|
158 |
-
|
159 |
-
|
|
|
|
|
160 |
elif subpart.startswith(">pm. "):
|
161 |
-
|
162 |
-
|
|
|
|
|
163 |
elif subpart.startswith(">w "):
|
164 |
self.type = 'B'
|
165 |
self.word = fix_text(subpart[3:])
|
@@ -168,6 +214,10 @@ class FR:
|
|
168 |
self.type = 'B'
|
169 |
self.word = fix_text(subpart[4:])
|
170 |
self.pseudoword = False
|
|
|
|
|
|
|
|
|
171 |
elif subpart.startswith("X"):
|
172 |
if hasattr(self, 'type'):
|
173 |
print(self.type, self.type == 'B')
|
@@ -176,49 +226,514 @@ class FR:
|
|
176 |
self.pseudoword = True
|
177 |
elif subpart == "OK":
|
178 |
self.type = 'E'
|
|
|
|
|
179 |
|
|
|
|
|
|
|
|
|
|
|
180 |
|
181 |
def __repr__(self):
|
182 |
parts = []
|
183 |
-
parts.append(f"type: {self.
|
184 |
parts.append(f"frame: {self.frame}")
|
185 |
-
if self.
|
186 |
-
parts.append(f"phone: {self.
|
187 |
if 'word' in self.__dict__:
|
188 |
parts.append(f"word: {self.word}")
|
189 |
if 'pm_type' in self.__dict__:
|
190 |
parts.append(f"pm_type: {self.pm_type}")
|
191 |
if 'pm' in self.__dict__:
|
192 |
parts.append(f"pm: {self.pm}")
|
193 |
-
|
194 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
195 |
|
196 |
|
197 |
class Mix():
|
198 |
-
def __init__(self, filepath: str):
|
199 |
self.fr = []
|
200 |
-
|
201 |
-
|
202 |
-
|
203 |
-
|
204 |
-
|
205 |
-
|
206 |
-
|
207 |
-
|
208 |
-
|
209 |
-
|
210 |
-
|
211 |
-
|
212 |
-
|
213 |
-
|
214 |
-
|
215 |
-
|
216 |
-
|
217 |
-
|
218 |
-
|
219 |
-
|
220 |
-
|
221 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
222 |
|
223 |
|
224 |
def smp_probe(filename: str) -> bool:
|
|
|
1 |
# coding=utf-8
|
2 |
# Copyright 2021 The TensorFlow Datasets Authors and the HuggingFace Datasets Authors.
|
3 |
+
# Copyright 2022, 2023 Jim O'Regan for Språkbanken Tal
|
4 |
#
|
5 |
# Licensed under the Apache License, Version 2.0 (the "License");
|
6 |
# you may not use this file except in compliance with the License.
|
|
|
20 |
from io import BytesIO
|
21 |
import os
|
22 |
import soundfile as sf
|
23 |
+
from collections import namedtuple
|
24 |
+
from copy import deepcopy
|
25 |
+
from difflib import SequenceMatcher
|
26 |
|
27 |
import datasets
|
28 |
from datasets.tasks import AutomaticSpeechRecognition
|
|
|
57 |
_URL = "http://www.speech.kth.se/waxholm/waxholm2.html"
|
58 |
|
59 |
|
60 |
+
class FRExpected(Exception):
|
61 |
+
"""Exception to raise when FR line was expected, but not read"""
|
62 |
+
def __init__(self, line):
|
63 |
+
msg = "Unknown line type (does not begin with 'FR'): "
|
64 |
+
super().__init__(msg + line)
|
65 |
+
|
66 |
+
|
67 |
class WaxholmDataset(datasets.GeneratorBasedBuilder):
|
68 |
"""Dataset script for Waxholm."""
|
69 |
|
|
|
78 |
{
|
79 |
"id": datasets.Value("string"),
|
80 |
"text": datasets.Value("string"),
|
81 |
+
"phonemes": datasets.Sequence(datasets.Value("string")),
|
82 |
"audio": datasets.Audio(sampling_rate=16_000)
|
83 |
}
|
84 |
)
|
|
|
90 |
homepage=_URL,
|
91 |
citation=_CITATION,
|
92 |
task_templates=[
|
93 |
+
AutomaticSpeechRecognition(audio_column="audio", transcription_column="text")
|
94 |
],
|
95 |
)
|
96 |
|
|
|
131 |
buffer = BytesIO()
|
132 |
sf.write(buffer, samples, sr, format="wav")
|
133 |
blank = Audio()
|
|
|
134 |
yield line, {
|
135 |
"id": line,
|
136 |
"text": mix.text,
|
137 |
+
"phonemes": mix.get_phoneme_list(),
|
138 |
+
"audio": {
|
139 |
+
"bytes": buffer.getvalue(),
|
140 |
+
"sampling_rate": sr,
|
141 |
+
}
|
142 |
}
|
143 |
|
144 |
|
145 |
def fix_text(text: str) -> str:
|
146 |
+
replacements = text.maketrans("{}|\\[]", "äåöÖÄÅ")
|
147 |
return text.translate(replacements)
|
148 |
|
149 |
|
150 |
+
Label = namedtuple('Label', ['start', 'end', 'label'])
|
151 |
+
|
152 |
+
|
153 |
class FR:
|
154 |
+
def __init__(self, text="", **kwargs): # C901
|
155 |
+
if text and text != "":
|
156 |
+
self.from_text(text)
|
157 |
+
else:
|
158 |
+
for arg in kwargs:
|
159 |
+
prms = ["pm", "pm_type", "type", "frame",
|
160 |
+
"seconds", "phone", "phone_type",
|
161 |
+
"word", "pseudoword"]
|
162 |
+
if arg in prms:
|
163 |
+
self.__dict__[arg] = kwargs[arg]
|
164 |
+
else:
|
165 |
+
print(f"Unrecognised argument: {arg}")
|
166 |
+
|
167 |
+
def from_text(self, text: str):
|
168 |
if not text.startswith("FR"):
|
169 |
+
raise FRExpected(text)
|
170 |
parts = [a.strip() for a in text.split("\t")]
|
171 |
self.frame = parts[0][2:].strip()
|
172 |
if parts[-1].strip().endswith(" sec"):
|
173 |
self.seconds = parts[-1].strip()[0:-4]
|
174 |
+
def split_phone(phone):
|
175 |
+
if phone.startswith("$#"):
|
176 |
+
phtype = 'I'
|
177 |
+
phone_type = fix_text(phone[0:2])
|
178 |
+
phone_out = fix_text(phone[2:])
|
179 |
+
elif phone.startswith("$") or phone.startswith("#"):
|
180 |
+
phtype = 'I'
|
181 |
+
phone_type = fix_text(phone[0:1])
|
182 |
+
phone_out = fix_text(phone[1:])
|
183 |
+
else:
|
184 |
+
print(phone)
|
185 |
+
return None
|
186 |
+
return {
|
187 |
+
"type": phtype,
|
188 |
+
"phone_type": phone_type,
|
189 |
+
"phone": phone_out
|
190 |
+
}
|
191 |
for subpart in parts[1:-1]:
|
192 |
+
subpart = subpart.strip()
|
193 |
+
if subpart.startswith("$#") or subpart.startswith("$") or subpart.startswith("#"):
|
194 |
+
phparts = split_phone(subpart)
|
195 |
+
if phparts is not None:
|
196 |
+
self.type = phparts['type']
|
197 |
+
self.phone_type = phparts['phone_type']
|
198 |
+
self.phone = phparts['phone']
|
|
|
|
|
|
|
|
|
|
|
199 |
elif subpart.startswith(">pm "):
|
200 |
+
phparts = split_phone(subpart[4:])
|
201 |
+
if phparts is not None:
|
202 |
+
self.pm_type = phparts['phone_type']
|
203 |
+
self.pm = phparts['phone']
|
204 |
elif subpart.startswith(">pm. "):
|
205 |
+
phparts = split_phone(subpart[5:])
|
206 |
+
if phparts is not None:
|
207 |
+
self.pm_type = phparts['phone_type']
|
208 |
+
self.pm = phparts['phone']
|
209 |
elif subpart.startswith(">w "):
|
210 |
self.type = 'B'
|
211 |
self.word = fix_text(subpart[3:])
|
|
|
214 |
self.type = 'B'
|
215 |
self.word = fix_text(subpart[4:])
|
216 |
self.pseudoword = False
|
217 |
+
elif subpart == "> XklickX" or subpart == "> XutandX":
|
218 |
+
self.type = 'B'
|
219 |
+
self.word = subpart[2:]
|
220 |
+
self.pseudoword = True
|
221 |
elif subpart.startswith("X"):
|
222 |
if hasattr(self, 'type'):
|
223 |
print(self.type, self.type == 'B')
|
|
|
226 |
self.pseudoword = True
|
227 |
elif subpart == "OK":
|
228 |
self.type = 'E'
|
229 |
+
elif subpart == "PROBLEMS":
|
230 |
+
self.type = 'E'
|
231 |
|
232 |
+
def get_type(self):
|
233 |
+
if "type" in self.__dict__:
|
234 |
+
return self.type
|
235 |
+
else:
|
236 |
+
return ""
|
237 |
|
238 |
def __repr__(self):
|
239 |
parts = []
|
240 |
+
parts.append(f"type: {self.get_type()}")
|
241 |
parts.append(f"frame: {self.frame}")
|
242 |
+
if self.get_type() != 'E':
|
243 |
+
parts.append(f"phone: {self.get_phone()}")
|
244 |
if 'word' in self.__dict__:
|
245 |
parts.append(f"word: {self.word}")
|
246 |
if 'pm_type' in self.__dict__:
|
247 |
parts.append(f"pm_type: {self.pm_type}")
|
248 |
if 'pm' in self.__dict__:
|
249 |
parts.append(f"pm: {self.pm}")
|
250 |
+
if 'seconds' in self.__dict__:
|
251 |
+
parts.append(f"sec: {self.seconds}")
|
252 |
+
return "FR(" + ", ".join(parts) + ")"
|
253 |
+
|
254 |
+
def fix_type(self):
|
255 |
+
if self.is_type("B") and self.get_word() == "":
|
256 |
+
self.pm_type = "$"
|
257 |
+
self.phone_type = "$"
|
258 |
+
self.type = "I"
|
259 |
+
|
260 |
+
def get_phone(self, fix_accents=True):
|
261 |
+
def fix_accents(phone, fix_accents=True):
|
262 |
+
if not fix_accents:
|
263 |
+
return phone
|
264 |
+
return phone.replace("'", "ˈ").replace('"', "ˌ")
|
265 |
+
if 'pm' in self.__dict__:
|
266 |
+
return fix_accents(self.pm, fix_accents)
|
267 |
+
elif 'phone' in self.__dict__:
|
268 |
+
return fix_accents(self.phone, fix_accents)
|
269 |
+
else:
|
270 |
+
return None
|
271 |
+
|
272 |
+
def is_silence_word(self, noise=False):
|
273 |
+
if 'word' in self.__dict__:
|
274 |
+
if not noise:
|
275 |
+
return self.word == "XX"
|
276 |
+
else:
|
277 |
+
return self.word.startswith("X") and self.word.endswith("X")
|
278 |
+
else:
|
279 |
+
return False
|
280 |
+
|
281 |
+
def is_type(self, type):
|
282 |
+
if "type" in self.__dict__:
|
283 |
+
return type == self.type
|
284 |
+
else:
|
285 |
+
return False
|
286 |
+
|
287 |
+
def has_seconds(self):
|
288 |
+
return "seconds" in self.__dict__
|
289 |
+
|
290 |
+
def get_seconds(self):
|
291 |
+
if not self.has_seconds() and "frame" in self.__dict__:
|
292 |
+
return int(self.frame) / 16000.0
|
293 |
+
else:
|
294 |
+
return self.seconds
|
295 |
+
|
296 |
+
def get_word(self):
|
297 |
+
if self.has_word():
|
298 |
+
return self.word
|
299 |
+
else:
|
300 |
+
return ""
|
301 |
+
|
302 |
+
def has_word(self):
|
303 |
+
return "word" in self.__dict__
|
304 |
+
|
305 |
+
def has_pseudoword(self):
|
306 |
+
return "pseudoword" in self.__dict__
|
307 |
+
|
308 |
+
|
309 |
+
def merge_frs(fr1, fr2, check_time=False):
|
310 |
+
"""
|
311 |
+
Merge FRS entries for plosives: by default, the
|
312 |
+
period of glottal closure and the burst are separately
|
313 |
+
annotated.
|
314 |
+
"""
|
315 |
+
if fr2.has_word():
|
316 |
+
return None
|
317 |
+
if check_time:
|
318 |
+
if fr1.get_seconds() != fr2.get_seconds():
|
319 |
+
return None
|
320 |
+
if _is_glottal_closure(fr1.get_phone(), fr2.get_phone()):
|
321 |
+
if not fr1.has_word():
|
322 |
+
return fr2
|
323 |
+
else:
|
324 |
+
word = None
|
325 |
+
if fr1.has_word():
|
326 |
+
word = fr1.word
|
327 |
+
pword = None
|
328 |
+
if fr1.has_pseudoword():
|
329 |
+
pword = fr1.pseudoword
|
330 |
+
return FR(pm=fr2.pm, pm_type=fr2.pm_type, type=fr2.type,
|
331 |
+
frame=fr2.frame, seconds=fr2.seconds, phone=fr2.phone,
|
332 |
+
phone_type=fr2.phone_type, word=word, pseudoword=pword)
|
333 |
+
|
334 |
+
|
335 |
+
SILS = {
|
336 |
+
"K": "k",
|
337 |
+
"G": "g",
|
338 |
+
"T": "t",
|
339 |
+
"D": "d",
|
340 |
+
"2T": "2t",
|
341 |
+
"2D": "2d",
|
342 |
+
"P": "p",
|
343 |
+
"B": "b"
|
344 |
+
}
|
345 |
+
def _is_glottal_closure(cur, next):
|
346 |
+
return cur in SILS and next == SILS[cur]
|
347 |
+
|
348 |
+
|
349 |
+
def _replace_glottal_closures(input):
|
350 |
+
input += ' '
|
351 |
+
for sil in SILS:
|
352 |
+
input = input.replace(f"{sil} {SILS[sil]} ", f"{SILS[sil]} ")
|
353 |
+
return input[:-1]
|
354 |
+
|
355 |
+
def _fix_duration_markers(input):
|
356 |
+
input += ' '
|
357 |
+
input = input.replace(":+ ", ": ")
|
358 |
+
return input[:-1]
|
359 |
|
360 |
|
361 |
class Mix():
|
362 |
+
def __init__(self, filepath: str, stringfile=None, fix_type=True):
|
363 |
self.fr = []
|
364 |
+
self.path = filepath
|
365 |
+
if stringfile is None:
|
366 |
+
with open(filepath) as inpf:
|
367 |
+
self.read_data(inpf.readlines())
|
368 |
+
else:
|
369 |
+
self.read_data(stringfile.split("\n"))
|
370 |
+
if fix_type:
|
371 |
+
for fr in self.fr:
|
372 |
+
fr.fix_type()
|
373 |
+
|
374 |
+
def read_data(self, inpf): # C901
|
375 |
+
"""read data from text of a .mix file"""
|
376 |
+
saw_text = False
|
377 |
+
saw_phoneme = False
|
378 |
+
saw_labels = False
|
379 |
+
for line in inpf:
|
380 |
+
if line.startswith("Waxholm dialog."):
|
381 |
+
self.filepath = line[15:].strip()
|
382 |
+
if line.startswith("TEXT:"):
|
383 |
+
saw_text = True
|
384 |
+
continue
|
385 |
+
if saw_text:
|
386 |
+
self.text = fix_text(line.strip())
|
387 |
+
saw_text = False
|
388 |
+
if line.startswith("PHONEME:"):
|
389 |
+
saw_phoneme = True
|
390 |
+
self.phoneme = fix_text(line[8:].strip())
|
391 |
+
if line[8:].strip().endswith("."):
|
392 |
+
saw_phoneme = False
|
393 |
+
continue
|
394 |
+
if saw_phoneme:
|
395 |
+
self.phoneme = fix_text(line.strip())
|
396 |
+
if line[8:].strip().endswith("."):
|
397 |
+
saw_phoneme = False
|
398 |
+
if line.startswith("FR "):
|
399 |
+
if saw_labels:
|
400 |
+
saw_labels = False
|
401 |
+
self.fr.append(FR(text=line))
|
402 |
+
if line.startswith("Labels: "):
|
403 |
+
self.labels = line[8:].strip()
|
404 |
+
saw_labels = True
|
405 |
+
if saw_labels and line.startswith(" "):
|
406 |
+
self.labels += line.strip()
|
407 |
+
|
408 |
+
def check_fr(self, verbose=False) -> bool:
|
409 |
+
"""
|
410 |
+
Simple sanity check: that there were FR lines,
|
411 |
+
and that the first was a start type, and
|
412 |
+
last was an end type.
|
413 |
+
"""
|
414 |
+
if 'fr' not in self.__dict__:
|
415 |
+
return False
|
416 |
+
if len(self.fr) == 0:
|
417 |
+
return False
|
418 |
+
start_end = self.fr[0].is_type("B") and self.fr[-1].is_type("E")
|
419 |
+
if verbose and not start_end:
|
420 |
+
if not self.fr[0].is_type("B"):
|
421 |
+
print(f"{self.path}: missing start type")
|
422 |
+
if not self.fr[-1].is_type("E"):
|
423 |
+
print(f"{self.path}: missing end type")
|
424 |
+
return start_end
|
425 |
+
|
426 |
+
def get_times(self, as_frames=False):
|
427 |
+
"""
|
428 |
+
get the times of each phoneme
|
429 |
+
"""
|
430 |
+
if not self.check_fr(verbose=True):
|
431 |
+
return []
|
432 |
+
if as_frames:
|
433 |
+
times = [int(x.frame) for x in self.fr]
|
434 |
+
else:
|
435 |
+
times = [float(x.seconds) for x in self.fr]
|
436 |
+
return times
|
437 |
+
|
438 |
+
def get_time_pairs(self, as_frames=False):
|
439 |
+
"""
|
440 |
+
get a list of tuples containing start and end times
|
441 |
+
By default, the times are in seconds; if `as_frames`
|
442 |
+
is set, the number of frames are returned instead.
|
443 |
+
"""
|
444 |
+
times = self.get_times(as_frames=as_frames)
|
445 |
+
starts = times[0:-1]
|
446 |
+
ends = times[1:]
|
447 |
+
return [x for x in zip(starts, ends)]
|
448 |
+
|
449 |
+
def prune_empty_presilences(self, verbose=False, include_noises=False):
|
450 |
+
"""
|
451 |
+
Remove empty silence markers (i.e., those with no distinct duration)
|
452 |
+
"""
|
453 |
+
self.orig_fr = deepcopy(self.fr)
|
454 |
+
i = 0
|
455 |
+
warned = False
|
456 |
+
def check_cur(cur, next):
|
457 |
+
if verbose and not cur.has_seconds():
|
458 |
+
print(f"Missing seconds: {self.path}\nLine: {cur}")
|
459 |
+
if verbose and not next.has_seconds():
|
460 |
+
print(f"Missing seconds: {self.path}\nLine: {next}")
|
461 |
+
return cur.get_seconds() == next.get_seconds() and cur.is_silence_word()
|
462 |
+
todel = []
|
463 |
+
while i < len(self.fr) - 1:
|
464 |
+
if check_cur(self.fr[i], self.fr[i + 1]):
|
465 |
+
if verbose:
|
466 |
+
if not warned:
|
467 |
+
warned = True
|
468 |
+
print(f"Empty silence in {self.path}:")
|
469 |
+
print(self.fr[i])
|
470 |
+
todel.append(i)
|
471 |
+
i += 1
|
472 |
+
if todel is not None and todel != []:
|
473 |
+
for chaff in todel.reverse():
|
474 |
+
del(self.fr[chaff])
|
475 |
+
|
476 |
+
def prune_empty_postsilences(self, verbose=False, include_noises=False):
|
477 |
+
"""
|
478 |
+
Remove empty silence markers (i.e., those with no distinct duration)
|
479 |
+
"""
|
480 |
+
if not "orig_fr" in self.__dict__:
|
481 |
+
self.orig_fr = deepcopy(self.fr)
|
482 |
+
i = 1
|
483 |
+
warned = False
|
484 |
+
def check_cur(cur, prev):
|
485 |
+
if verbose and not cur.has_seconds():
|
486 |
+
print(f"Missing seconds: {self.path}\nLine: {cur}")
|
487 |
+
if verbose and not prev.has_seconds():
|
488 |
+
print(f"Missing seconds: {self.path}\nLine: {prev}")
|
489 |
+
return cur.get_seconds() == prev.get_seconds() and cur.is_silence_word()
|
490 |
+
todel = []
|
491 |
+
while i < len(self.fr):
|
492 |
+
if check_cur(self.fr[i], self.fr[i - 1]):
|
493 |
+
if verbose:
|
494 |
+
if not warned:
|
495 |
+
warned = True
|
496 |
+
print(f"Empty silence in {self.path}:")
|
497 |
+
print(self.fr[i])
|
498 |
+
todel.append(i)
|
499 |
+
i += 1
|
500 |
+
if todel is not None and todel != []:
|
501 |
+
for chaff in todel.reverse():
|
502 |
+
del(self.fr[chaff])
|
503 |
+
|
504 |
+
def prune_empty_segments(self, verbose=False):
|
505 |
+
"""
|
506 |
+
Remove empty segments (i.e., those with no distinct duration)
|
507 |
+
"""
|
508 |
+
if not "orig_fr" in self.__dict__:
|
509 |
+
self.orig_fr = deepcopy(self.fr)
|
510 |
+
times = self.get_time_pairs(as_frames=True)
|
511 |
+
if len(times) != (len(self.fr) - 1):
|
512 |
+
print("Uh oh: time pairs and items don't match")
|
513 |
+
else:
|
514 |
+
keep = []
|
515 |
+
for fr in zip(self.fr[:-1], times):
|
516 |
+
cur_time = fr[1]
|
517 |
+
if cur_time[0] == cur_time[1]:
|
518 |
+
if verbose:
|
519 |
+
print(f"Empty segment {fr[0].get_phone()} ({cur_time[0]} --> {cur_time[1]})")
|
520 |
+
else:
|
521 |
+
keep.append(fr[0])
|
522 |
+
keep.append(self.fr[-1])
|
523 |
+
self.fr = keep
|
524 |
+
|
525 |
+
def prune_empty_silences(self, verbose = False):
|
526 |
+
self.prune_empty_presilences(verbose)
|
527 |
+
self.prune_empty_postsilences(verbose)
|
528 |
+
|
529 |
+
def merge_plosives(self, verbose=False):
|
530 |
+
"""
|
531 |
+
Merge plosives in FRs
|
532 |
+
(in Waxholm, as in TIMIT, the silence before the burst and the burst
|
533 |
+
are annotated separately).
|
534 |
+
"""
|
535 |
+
if not "orig_fr" in self.__dict__:
|
536 |
+
self.orig_fr = deepcopy(self.fr)
|
537 |
+
tmp = []
|
538 |
+
i = 0
|
539 |
+
while i < len(self.fr)-1:
|
540 |
+
merged = merge_frs(self.fr[i], self.fr[i+1])
|
541 |
+
if merged is not None:
|
542 |
+
if verbose:
|
543 |
+
print(f"Merging {self.fr[i]} and {self.fr[i+1]}")
|
544 |
+
i += 1
|
545 |
+
tmp.append(merged)
|
546 |
+
else:
|
547 |
+
tmp.append(self.fr[i])
|
548 |
+
i += 1
|
549 |
+
tmp.append(self.fr[-1])
|
550 |
+
self.fr = tmp
|
551 |
+
|
552 |
+
def get_phone_label_tuples(self, as_frames=False, fix_accents=True):
|
553 |
+
times = self.get_time_pairs(as_frames=as_frames)
|
554 |
+
if self.check_fr():
|
555 |
+
labels = [fr.get_phone(fix_accents) for fr in self.fr[0:-1]]
|
556 |
+
else:
|
557 |
+
labels = []
|
558 |
+
if len(times) == len(labels):
|
559 |
+
out = []
|
560 |
+
for z in zip(times, labels):
|
561 |
+
out.append((z[0][0], z[0][1], z[1]))
|
562 |
+
return out
|
563 |
+
else:
|
564 |
+
return []
|
565 |
+
|
566 |
+
def get_merged_plosives(self, noop=False, prune_empty=True):
|
567 |
+
"""
|
568 |
+
Returns a list of phones with plosives merged
|
569 |
+
(in Waxholm, as in TIMIT, the silence before the burst and the burst
|
570 |
+
are annotated separately).
|
571 |
+
If `noop` is True, it simply returns the output of `prune_empty_labels()`
|
572 |
+
"""
|
573 |
+
if noop:
|
574 |
+
if not prune_empty:
|
575 |
+
print("Warning: not valid to set noop to True and prune_empty to false")
|
576 |
+
print("Ignoring prune_empty")
|
577 |
+
return self.prune_empty_labels()
|
578 |
+
i = 0
|
579 |
+
out = []
|
580 |
+
if prune_empty:
|
581 |
+
labels = self.prune_empty_labels()
|
582 |
+
else:
|
583 |
+
labels = self.get_phone_label_tuples()
|
584 |
+
while i < len(labels)-1:
|
585 |
+
cur = labels[i]
|
586 |
+
next = labels[i+1]
|
587 |
+
if _is_glottal_closure(cur[2], next[2]):
|
588 |
+
tmp = Label(start = cur[0], end = next[1], label = next[2])
|
589 |
+
out.append(tmp)
|
590 |
+
i += 2
|
591 |
+
else:
|
592 |
+
tmp = Label(start = cur[0], end = cur[1], label = cur[2])
|
593 |
+
out.append(tmp)
|
594 |
+
i += 1
|
595 |
+
return out
|
596 |
+
|
597 |
+
def get_word_label_tuples(self, verbose=True):
|
598 |
+
times = self.get_time_pairs()
|
599 |
+
if len(times) == len(self.fr[0:-1]):
|
600 |
+
out = []
|
601 |
+
labels_raw = [x for x in zip(times, self.fr[0:-1])]
|
602 |
+
i = 0
|
603 |
+
cur = None
|
604 |
+
while i < len(labels_raw) - 1:
|
605 |
+
if labels_raw[i][1].is_type("B"):
|
606 |
+
if cur is not None:
|
607 |
+
out.append(cur)
|
608 |
+
if labels_raw[i+1][1].is_type("B"):
|
609 |
+
if verbose and labels_raw[i][1].get_word() == "":
|
610 |
+
print("Expected word", labels_raw[i][1])
|
611 |
+
out.append((labels_raw[i][0][0], labels_raw[i][0][1], labels_raw[i][1].get_word()))
|
612 |
+
cur = None
|
613 |
+
i += 1
|
614 |
+
continue
|
615 |
+
else:
|
616 |
+
if verbose and labels_raw[i][1].get_word() == "":
|
617 |
+
print("Expected word", labels_raw[i][1])
|
618 |
+
cur = (labels_raw[i][0][0], labels_raw[i][0][1], labels_raw[i][1].get_word())
|
619 |
+
if labels_raw[i+1][1].is_type("B"):
|
620 |
+
if cur is not None:
|
621 |
+
cur = (cur[0], labels_raw[i][0][1], cur[2])
|
622 |
+
i += 1
|
623 |
+
out.append(cur)
|
624 |
+
return out
|
625 |
+
else:
|
626 |
+
return []
|
627 |
+
|
628 |
+
def get_dictionary(self, fix_accents=True):
|
629 |
+
"""
|
630 |
+
Get pronunciation dictionary entries from the .mix file.
|
631 |
+
These entries are based on the corrected pronunciations; for
|
632 |
+
the lexical pronunciations, use the `phoneme` property.
|
633 |
+
"""
|
634 |
+
output = {}
|
635 |
+
current_phones = []
|
636 |
+
prev_word = ''
|
637 |
+
|
638 |
+
for fr in self.fr:
|
639 |
+
if 'word' in fr.__dict__:
|
640 |
+
phone = fr.get_phone(fix_accents)
|
641 |
+
if prev_word != "":
|
642 |
+
if prev_word not in output:
|
643 |
+
output[prev_word] = []
|
644 |
+
output[prev_word].append(current_phones.copy())
|
645 |
+
current_phones.clear()
|
646 |
+
prev_word = fr.word
|
647 |
+
current_phones.append(phone)
|
648 |
+
elif fr.is_type("I"):
|
649 |
+
phone = fr.get_phone(fix_accents)
|
650 |
+
current_phones.append(phone)
|
651 |
+
else:
|
652 |
+
if prev_word not in output:
|
653 |
+
output[prev_word] = []
|
654 |
+
output[prev_word].append(current_phones.copy())
|
655 |
+
return output
|
656 |
+
|
657 |
+
def get_dictionary_list(self, fix_accents=True):
|
658 |
+
"""
|
659 |
+
Get pronunciation dictionary entries from the .mix file.
|
660 |
+
These entries are based on the corrected pronunciations; for
|
661 |
+
the lexical pronunciations, use the `phoneme` property.
|
662 |
+
This version creates a list of tuples (word, phones) that
|
663 |
+
preserves the order of the entries.
|
664 |
+
"""
|
665 |
+
output = []
|
666 |
+
current_phones = []
|
667 |
+
prev_word = ''
|
668 |
+
|
669 |
+
for fr in self.fr:
|
670 |
+
if 'word' in fr.__dict__:
|
671 |
+
phone = fr.get_phone(fix_accents)
|
672 |
+
if prev_word != "":
|
673 |
+
output.append((prev_word, " ".join(current_phones)))
|
674 |
+
current_phones.clear()
|
675 |
+
prev_word = fr.word
|
676 |
+
current_phones.append(phone)
|
677 |
+
elif fr.is_type("I"):
|
678 |
+
phone = fr.get_phone(fix_accents)
|
679 |
+
current_phones.append(phone)
|
680 |
+
else:
|
681 |
+
output.append((prev_word, " ".join(current_phones)))
|
682 |
+
return output
|
683 |
+
|
684 |
+
def get_phoneme_string(self, insert_pauses=True, fix_accents=True):
|
685 |
+
"""
|
686 |
+
Get an opinionated phoneme string
|
687 |
+
|
688 |
+
Args:
|
689 |
+
insert_pauses (bool, optional): Insert pauses between words. Defaults to True.
|
690 |
+
fix_accents (bool, optional): IPA-ify accents. Defaults to True.
|
691 |
+
"""
|
692 |
+
dict_list = self.get_dictionary_list(fix_accents)
|
693 |
+
skip = ['p:', '.']
|
694 |
+
if insert_pauses:
|
695 |
+
phone_strings = [x[1] for x in dict_list if x[1] not in skip]
|
696 |
+
joined = ' p: '.join(phone_strings)
|
697 |
+
else:
|
698 |
+
phone_strings = [x[1] for x in dict_list if x[1] != "."]
|
699 |
+
joined = ' '.join(phone_strings)
|
700 |
+
joined = _replace_glottal_closures(joined)
|
701 |
+
joined = _fix_duration_markers(joined)
|
702 |
+
return joined
|
703 |
+
|
704 |
+
def get_phoneme_list(self, insert_pauses=True, fix_accents=True):
|
705 |
+
return self.get_phoneme_string(insert_pauses, fix_accents).split(' ')
|
706 |
+
|
707 |
+
def get_compare_dictionary(self, fix_accents=True, merge_plosives=True, only_changed=True):
|
708 |
+
"""
|
709 |
+
Get pronunciation dictionary for comparision: i.e., where there is a difference
|
710 |
+
between the canonical pronunciation and what was spoken
|
711 |
+
"""
|
712 |
+
if merge_plosives:
|
713 |
+
self.merge_plosives()
|
714 |
+
orig = self.get_dictionary_list(fix_accents)
|
715 |
+
self.prune_empty_segments()
|
716 |
+
new = self.get_dictionary_list(fix_accents)
|
717 |
+
if len(orig) != len(new):
|
718 |
+
words_orig = [w[0] for w in orig]
|
719 |
+
words_new = [w[0] for w in new]
|
720 |
+
skippables = []
|
721 |
+
for tag, i, j, _, _ in SequenceMatcher(None, words_orig, words_new).get_opcodes():
|
722 |
+
if tag in ('delete', 'replace'):
|
723 |
+
skippables += [a for a in range(i, j)]
|
724 |
+
for c in skippables.reverse():
|
725 |
+
del(orig[c])
|
726 |
+
out = []
|
727 |
+
i = 0
|
728 |
+
while i < len(orig):
|
729 |
+
if orig[i][0] == new[i][0]:
|
730 |
+
if orig[i][1] == new[i][1]:
|
731 |
+
if not only_changed:
|
732 |
+
out.append(orig)
|
733 |
+
else:
|
734 |
+
out.append((orig[i][0], orig[i][1], new[i][1]))
|
735 |
+
i += 1
|
736 |
+
return out
|
737 |
|
738 |
|
739 |
def smp_probe(filename: str) -> bool:
|