Upload 2 files
Browse files- ctcalign.py +166 -169
- graph.py +3 -5
ctcalign.py
CHANGED
@@ -4,108 +4,98 @@ import numpy as np
|
|
4 |
from dataclasses import dataclass
|
5 |
|
6 |
|
|
|
|
|
7 |
|
|
|
|
|
|
|
|
|
8 |
|
9 |
-
|
10 |
-
#
|
11 |
-
|
12 |
-
return fr/50
|
13 |
|
14 |
-
|
15 |
-
|
16 |
-
#
|
17 |
-
#------------------------------------------
|
18 |
|
19 |
-
# important to know for CTC decoding - potentially language/model dependent
|
20 |
-
#model_word_separator = '|'
|
21 |
-
#model_blank_token = '[PAD]'
|
22 |
-
#is_MODEL_PATH="../models/LVL/wav2vec2-large-xlsr-53-icelandic-ep10-1000h"
|
23 |
|
|
|
|
|
|
|
|
|
24 |
|
25 |
|
26 |
-
|
27 |
-
|
28 |
-
def
|
29 |
-
|
30 |
-
|
31 |
-
|
32 |
-
self.model = Wav2Vec2ForCTC.from_pretrained(model_path)#.to(self.device)
|
33 |
-
self.processor = Wav2Vec2Processor.from_pretrained(model_path)
|
34 |
-
|
35 |
-
# build labels dict from a processor where it is not directly accessible
|
36 |
-
max_labels = 100 # any reasonable number higher than vocab + extra + special tokens in any language used
|
37 |
-
ixs = sorted(list(range(max_labels)),reverse=True)
|
38 |
-
self.labels_dict = {self.processor.tokenizer.decode(n) or model_word_separator:n for n in ixs}
|
39 |
-
|
40 |
-
self.blank_id = self.labels_dict[model_blank_token]
|
41 |
-
self.model_word_separator = model_word_separator
|
42 |
|
43 |
|
|
|
|
|
|
|
|
|
|
|
44 |
|
45 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
46 |
|
47 |
-
#------------------------------------------
|
48 |
-
# forced alignment with ctc decoder
|
49 |
-
# based on implementation of
|
50 |
-
# https://pytorch.org/audio/main/tutorials/forced_alignment_tutorial.html
|
51 |
-
#------------------------------------------
|
52 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
53 |
|
54 |
-
# return the label class probability of each audio frame
|
55 |
-
# wav is the wav data already read in, NOT the file path.
|
56 |
-
def get_frame_probs(wav,aligner):
|
57 |
-
with torch.inference_mode(): # similar to with torch.no_grad():
|
58 |
-
input_values = aligner.processor(wav,sampling_rate=16000).input_values[0]
|
59 |
-
input_values = torch.tensor(input_values).unsqueeze(0)#, device=aligner.device).unsqueeze(0)
|
60 |
-
emits = aligner.model(input_values).logits
|
61 |
-
emits = torch.log_softmax(emits, dim=-1)
|
62 |
-
return emits[0].cpu().detach()
|
63 |
|
64 |
|
65 |
-
|
|
|
|
|
|
|
|
|
66 |
|
67 |
-
|
68 |
-
|
69 |
-
|
70 |
-
|
71 |
-
|
72 |
-
|
73 |
-
|
74 |
-
|
75 |
-
|
76 |
-
|
77 |
-
|
78 |
-
|
79 |
-
|
80 |
-
|
81 |
-
return trellis
|
82 |
-
|
83 |
-
|
84 |
-
|
85 |
-
@dataclass
|
86 |
-
class Point:
|
87 |
-
token_index: int
|
88 |
-
time_index: int
|
89 |
-
score: float
|
90 |
|
91 |
-
@dataclass
|
92 |
-
class Segment:
|
93 |
-
label: str
|
94 |
-
start: int
|
95 |
-
end: int
|
96 |
-
score: float
|
97 |
-
|
98 |
-
@property
|
99 |
-
def mfaform(self):
|
100 |
-
return f"{f2s(self.start)},{f2s(self.end)},{self.label}"
|
101 |
-
|
102 |
-
@property
|
103 |
-
def length(self):
|
104 |
-
return self.end - self.start
|
105 |
|
106 |
|
107 |
-
|
108 |
-
def backtrack(trellis, emission, tokens, blank_id):
|
109 |
# Note:
|
110 |
# j and t are indices for trellis, which has extra dimensions
|
111 |
# for time and tokens at the beginning.
|
@@ -113,112 +103,119 @@ def backtrack(trellis, emission, tokens, blank_id):
|
|
113 |
# the corresponding index in emission is `T-1`.
|
114 |
# Similarly, when referring to token index `J` in trellis,
|
115 |
# the corresponding index in transcript is `J-1`.
|
116 |
-
|
117 |
-
|
118 |
|
119 |
-
|
120 |
-
|
121 |
# 1. Figure out if the current position was stay or change
|
122 |
# `emission[J-1]` is the emission at time frame `J` of trellis dimension.
|
123 |
# Score for token staying the same from time frame J-1 to T.
|
124 |
-
|
125 |
# Score for token changing from C-1 at T-1 to J at T.
|
126 |
-
|
127 |
|
128 |
# 2. Store the path with frame-wise probability.
|
129 |
-
|
130 |
# Return token index and time index in non-trellis coordinate.
|
131 |
-
|
132 |
-
|
133 |
-
|
134 |
-
|
135 |
-
|
136 |
-
|
137 |
-
|
138 |
-
else:
|
139 |
-
raise ValueError("Failed to align")
|
140 |
-
return path[::-1]
|
141 |
-
|
142 |
-
|
143 |
-
def merge_repeats(path,transcript):
|
144 |
-
i1, i2 = 0, 0
|
145 |
-
segments = []
|
146 |
-
while i1 < len(path):
|
147 |
-
while i2 < len(path) and path[i1].token_index == path[i2].token_index: # while both path steps point to the same token index
|
148 |
-
i2 += 1
|
149 |
-
score = sum(path[k].score for k in range(i1, i2)) / (i2 - i1)
|
150 |
-
segments.append( # when i2 finally switches to a different token,
|
151 |
-
Segment(
|
152 |
-
transcript[path[i1].token_index],# to the list of segments, append the token from i1
|
153 |
-
path[i1].time_index, # time of the first path-point of that token
|
154 |
-
path[i2 - 1].time_index + 1, # time of the final path-point for that token.
|
155 |
-
score,
|
156 |
-
)
|
157 |
-
)
|
158 |
-
i1 = i2
|
159 |
-
return segments
|
160 |
-
|
161 |
-
|
162 |
-
|
163 |
-
def merge_words(segments, separator):
|
164 |
-
words = []
|
165 |
-
i1, i2 = 0, 0
|
166 |
-
while i1 < len(segments):
|
167 |
-
if i2 >= len(segments) or segments[i2].label == separator:
|
168 |
-
if i1 != i2:
|
169 |
-
segs = segments[i1:i2]
|
170 |
-
word = "".join([seg.label for seg in segs])
|
171 |
-
score = sum(seg.score * seg.length for seg in segs) / sum(seg.length for seg in segs)
|
172 |
-
words.append(Segment(word, segments[i1].start, segments[i2 - 1].end, score))
|
173 |
-
i1 = i2 + 1
|
174 |
-
i2 = i1
|
175 |
else:
|
176 |
-
|
177 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
178 |
|
179 |
|
180 |
|
181 |
-
|
182 |
-
|
183 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
184 |
|
185 |
|
186 |
-
# generate mfa format for character (phone) and word alignments
|
187 |
-
# skip the word separator as it is not a phone
|
188 |
-
def mfalike(chars,wds,wsep):
|
189 |
-
hed = ['Begin,End,Label,Type,Speaker\n']
|
190 |
-
wlines = [f'{w.mfaform},words,000\n' for w in wds]
|
191 |
-
slines = [f'{ch.mfaform},phones,000\n' for ch in chars if ch.label != wsep]
|
192 |
-
return (''.join(hed+wlines+slines))
|
193 |
-
|
194 |
-
# generate basic exportable list format for character OR word alignments
|
195 |
-
# skip the word separator as it is not a phone
|
196 |
-
def basic(segs,wsep="|"):
|
197 |
-
return [[s.label,f2s(s.start),f2s(s.end)] for s in segs if s.label != wsep]
|
198 |
-
|
199 |
|
200 |
-
# needs pad labels added to correctly time first segment
|
201 |
-
# and therefore add word sep character as placeholder in transcript
|
202 |
-
def prep_transcript(xcp, aligner):
|
203 |
-
xcp = xcp.replace(' ', aligner.model_word_separator)
|
204 |
-
label_ids = [aligner.labels_dict[c] for c in xcp]
|
205 |
-
label_ids = [aligner.blank_id] + label_ids + [aligner.blank_id]
|
206 |
-
xcp = f'{ aligner.model_word_separator}{xcp}{aligner.model_word_separator}'
|
207 |
-
return xcp,label_ids
|
208 |
|
|
|
|
|
|
|
209 |
|
210 |
|
211 |
-
|
212 |
-
|
213 |
-
|
214 |
-
|
215 |
-
|
|
|
|
|
216 |
|
217 |
-
|
218 |
-
|
|
|
|
|
219 |
|
220 |
-
|
221 |
-
|
222 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
223 |
|
224 |
|
|
|
4 |
from dataclasses import dataclass
|
5 |
|
6 |
|
7 |
+
|
8 |
+
def aligner(model_path,model_word_separator = '|', model_blank_token = '[PAD]'):
|
9 |
|
10 |
+
# build labels dict from a processor where it is not directly accessible
|
11 |
+
def get_processor_labels(processor,word_sep,max_labels=100):
|
12 |
+
ixs = sorted(list(range(max_labels)),reverse=True)
|
13 |
+
return {processor.tokenizer.decode(n) or word_sep:n for n in ixs}
|
14 |
|
15 |
+
#------------------------------------------
|
16 |
+
# setup wav2vec2
|
17 |
+
#------------------------------------------
|
|
|
18 |
|
19 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
20 |
+
torch.random.manual_seed(0)
|
21 |
+
max_labels = 100 # any reasonable number higher than vocab + extra + special tokens in any language used
|
|
|
22 |
|
|
|
|
|
|
|
|
|
23 |
|
24 |
+
model = Wav2Vec2ForCTC.from_pretrained(model_path).to(device)
|
25 |
+
processor = Wav2Vec2Processor.from_pretrained(model_path)
|
26 |
+
labels_dict = get_processor_labels(processor,model_word_separator)
|
27 |
+
blank_id = labels_dict[model_blank_token]
|
28 |
|
29 |
|
30 |
+
#convert frame-numbers to timestamps in seconds
|
31 |
+
# w2v2 step size is about 20ms, or 50 frames per second
|
32 |
+
def f2s(fr):
|
33 |
+
return fr/50
|
34 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
35 |
|
36 |
|
37 |
+
#------------------------------------------
|
38 |
+
# forced alignment with ctc decoder
|
39 |
+
# based on implementation of
|
40 |
+
# https://pytorch.org/audio/main/tutorials/forced_alignment_tutorial.html
|
41 |
+
#------------------------------------------
|
42 |
|
43 |
|
44 |
+
# return the label class probability of each audio frame
|
45 |
+
# wav is the wav data already read in, NOT the file path.
|
46 |
+
def get_frame_probs(wav):
|
47 |
+
with torch.inference_mode(): # similar to with torch.no_grad():
|
48 |
+
input_values = processor(wav,sampling_rate=16000).input_values[0]
|
49 |
+
input_values = torch.tensor(input_values, device=device).unsqueeze(0)
|
50 |
+
emits = model(input_values).logits
|
51 |
+
emits = torch.log_softmax(emits, dim=-1)
|
52 |
+
return emits[0].cpu().detach()
|
53 |
|
|
|
|
|
|
|
|
|
|
|
54 |
|
55 |
+
def get_trellis(emission, tokens, blank_id):
|
56 |
+
|
57 |
+
num_frame = emission.size(0)
|
58 |
+
num_tokens = len(tokens)
|
59 |
+
trellis = torch.empty((num_frame + 1, num_tokens + 1))
|
60 |
+
trellis[0, 0] = 0
|
61 |
+
trellis[1:, 0] = torch.cumsum(emission[:, 0], 0) # len of this slice of trellis is len of audio frames)
|
62 |
+
trellis[0, -num_tokens:] = -float("inf") # len of this slice of trellis is len of transcript tokens
|
63 |
+
trellis[-num_tokens:, 0] = float("inf")
|
64 |
+
for t in range(num_frame):
|
65 |
+
trellis[t + 1, 1:] = torch.maximum(
|
66 |
+
# Score for staying at the same token
|
67 |
+
trellis[t, 1:] + emission[t, blank_id],
|
68 |
+
# Score for changing to the next token
|
69 |
+
trellis[t, :-1] + emission[t, tokens],
|
70 |
+
)
|
71 |
+
return trellis
|
72 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
73 |
|
74 |
|
75 |
+
@dataclass
|
76 |
+
class Point:
|
77 |
+
token_index: int
|
78 |
+
time_index: int
|
79 |
+
score: float
|
80 |
|
81 |
+
@dataclass
|
82 |
+
class Segment:
|
83 |
+
label: str
|
84 |
+
start: int
|
85 |
+
end: int
|
86 |
+
score: float
|
87 |
+
|
88 |
+
@property
|
89 |
+
def mfaform(self):
|
90 |
+
return f"{f2s(self.start)},{f2s(self.end)},{self.label}"
|
91 |
+
|
92 |
+
@property
|
93 |
+
def length(self):
|
94 |
+
return self.end - self.start
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
95 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
96 |
|
97 |
|
98 |
+
def backtrack(trellis, emission, tokens, blank_id):
|
|
|
99 |
# Note:
|
100 |
# j and t are indices for trellis, which has extra dimensions
|
101 |
# for time and tokens at the beginning.
|
|
|
103 |
# the corresponding index in emission is `T-1`.
|
104 |
# Similarly, when referring to token index `J` in trellis,
|
105 |
# the corresponding index in transcript is `J-1`.
|
106 |
+
j = trellis.size(1) - 1
|
107 |
+
t_start = torch.argmax(trellis[:, j]).item()
|
108 |
|
109 |
+
path = []
|
110 |
+
for t in range(t_start, 0, -1):
|
111 |
# 1. Figure out if the current position was stay or change
|
112 |
# `emission[J-1]` is the emission at time frame `J` of trellis dimension.
|
113 |
# Score for token staying the same from time frame J-1 to T.
|
114 |
+
stayed = trellis[t - 1, j] + emission[t - 1, blank_id]
|
115 |
# Score for token changing from C-1 at T-1 to J at T.
|
116 |
+
changed = trellis[t - 1, j - 1] + emission[t - 1, tokens[j - 1]]
|
117 |
|
118 |
# 2. Store the path with frame-wise probability.
|
119 |
+
prob = emission[t - 1, tokens[j - 1] if changed > stayed else 0].exp().item()
|
120 |
# Return token index and time index in non-trellis coordinate.
|
121 |
+
path.append(Point(j - 1, t - 1, prob))
|
122 |
+
|
123 |
+
# 3. Update the token
|
124 |
+
if changed > stayed:
|
125 |
+
j -= 1
|
126 |
+
if j == 0:
|
127 |
+
break
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
128 |
else:
|
129 |
+
raise ValueError("Failed to align")
|
130 |
+
return path[::-1]
|
131 |
+
|
132 |
+
|
133 |
+
def merge_repeats(path,transcript):
|
134 |
+
i1, i2 = 0, 0
|
135 |
+
segments = []
|
136 |
+
while i1 < len(path):
|
137 |
+
while i2 < len(path) and path[i1].token_index == path[i2].token_index: # while both path steps point to the same token index
|
138 |
+
i2 += 1
|
139 |
+
score = sum(path[k].score for k in range(i1, i2)) / (i2 - i1)
|
140 |
+
segments.append( # when i2 finally switches to a different token,
|
141 |
+
Segment(
|
142 |
+
transcript[path[i1].token_index],# to the list of segments, append the token from i1
|
143 |
+
path[i1].time_index, # time of the first path-point of that token
|
144 |
+
path[i2 - 1].time_index + 1, # time of the final path-point for that token.
|
145 |
+
score,
|
146 |
+
)
|
147 |
+
)
|
148 |
+
i1 = i2
|
149 |
+
return segments
|
150 |
|
151 |
|
152 |
|
153 |
+
def merge_words(segments, separator):
|
154 |
+
words = []
|
155 |
+
i1, i2 = 0, 0
|
156 |
+
while i1 < len(segments):
|
157 |
+
if i2 >= len(segments) or segments[i2].label == separator:
|
158 |
+
if i1 != i2:
|
159 |
+
segs = segments[i1:i2]
|
160 |
+
word = "".join([seg.label for seg in segs])
|
161 |
+
score = sum(seg.score * seg.length for seg in segs) / sum(seg.length for seg in segs)
|
162 |
+
words.append(Segment(word, segments[i1].start, segments[i2 - 1].end, score))
|
163 |
+
i1 = i2 + 1
|
164 |
+
i2 = i1
|
165 |
+
else:
|
166 |
+
i2 += 1
|
167 |
+
return words
|
168 |
|
169 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
170 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
171 |
|
172 |
+
#------------------------------------------
|
173 |
+
# handle etc.
|
174 |
+
#------------------------------------------
|
175 |
|
176 |
|
177 |
+
# generate mfa format for character (phone) and word alignments
|
178 |
+
# skip the word separator as it is not a phone
|
179 |
+
def mfalike(chars,wds,wsep):
|
180 |
+
hed = ['Begin,End,Label,Type,Speaker\n']
|
181 |
+
wlines = [f'{w.mfaform},words,000\n' for w in wds]
|
182 |
+
slines = [f'{ch.mfaform},phones,000\n' for ch in chars if ch.label != wsep]
|
183 |
+
return (''.join(hed+wlines+slines))
|
184 |
|
185 |
+
# generate basic exportable list format for character OR word alignments
|
186 |
+
# skip the word separator as it is not a phone
|
187 |
+
def basic(segs,wsep="|"):
|
188 |
+
return [[s.label,f2s(s.start),f2s(s.end)] for s in segs if s.label != wsep]
|
189 |
|
190 |
+
|
191 |
+
# needs pad labels added to correctly time first segment
|
192 |
+
# and therefore add word sep character as placeholder in transcript
|
193 |
+
def prep_transcript(xcp):
|
194 |
+
xcp = xcp.replace(' ',model_word_separator)
|
195 |
+
label_ids = [labels_dict[c] for c in xcp]
|
196 |
+
label_ids = [blank_id] + label_ids + [blank_id]
|
197 |
+
xcp = f'{model_word_separator}{xcp}{model_word_separator}'
|
198 |
+
return xcp,label_ids
|
199 |
+
|
200 |
+
|
201 |
+
|
202 |
+
def _align(wav_data,transcript):
|
203 |
+
|
204 |
+
norm_transcript,rec_label_ids = prep_transcript(transcript)
|
205 |
+
emit = get_frame_probs(wav_data)
|
206 |
+
trellis = get_trellis(emit, rec_label_ids, blank_id)
|
207 |
+
path = backtrack(trellis, emit, rec_label_ids, blank_id)
|
208 |
+
|
209 |
+
segments = merge_repeats(path,norm_transcript)
|
210 |
+
words = merge_words(segments, model_word_separator)
|
211 |
+
|
212 |
+
#segments = [s for s in segments if s[0] != model_word_separator]
|
213 |
+
#return mfalike(segments,words,model_word_separator)
|
214 |
+
return basic(words,model_word_separator), basic(segments,model_word_separator)
|
215 |
+
|
216 |
+
return _align
|
217 |
+
|
218 |
+
|
219 |
+
|
220 |
|
221 |
|
graph.py
CHANGED
@@ -4,7 +4,6 @@ from scipy import signal
|
|
4 |
import librosa
|
5 |
import subprocess
|
6 |
import matplotlib.pyplot as plt
|
7 |
-
import ctcalign
|
8 |
|
9 |
|
10 |
|
@@ -44,13 +43,12 @@ def get_pitch_tracks(wav_path):
|
|
44 |
# transcript could be from a corpus with the wav file,
|
45 |
# input by the user,
|
46 |
# or from a previous speech recognition process
|
47 |
-
def align_and_graph(wav_path, transcript,
|
48 |
|
49 |
# fetch data
|
50 |
-
#f0_data = get_pitch_tracks(wav_path)
|
51 |
speech = readwav(wav_path)
|
52 |
-
w_align, seg_align =
|
53 |
-
|
54 |
|
55 |
# set up the graph shape
|
56 |
rec_start = w_align[0][1]
|
|
|
4 |
import librosa
|
5 |
import subprocess
|
6 |
import matplotlib.pyplot as plt
|
|
|
7 |
|
8 |
|
9 |
|
|
|
43 |
# transcript could be from a corpus with the wav file,
|
44 |
# input by the user,
|
45 |
# or from a previous speech recognition process
|
46 |
+
def align_and_graph(wav_path, transcript, aligner_function):
|
47 |
|
48 |
# fetch data
|
|
|
49 |
speech = readwav(wav_path)
|
50 |
+
w_align, seg_align = aligner_function(speech,normalise_transcript(transcript))
|
51 |
+
|
52 |
|
53 |
# set up the graph shape
|
54 |
rec_start = w_align[0][1]
|