Update README.md
Browse files
README.md
CHANGED
@@ -14,7 +14,7 @@ base_model:
|
|
14 |
# Wav2Vec2Bert Audio frame classifier for prosodic unit detection
|
15 |
|
16 |
This model predicts prosodic units on speech.
|
17 |
-
For each 20ms frame the model predicts
|
18 |
this frame or not.
|
19 |
|
20 |
|
@@ -40,12 +40,237 @@ This is the model card of a 🤗 transformers model that has been pushed on the
|
|
40 |
|
41 |
## Uses
|
42 |
|
|
|
43 |
|
|
|
|
|
|
|
44 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
45 |
## Bias, Risks, and Limitations
|
46 |
|
47 |
## Training Details
|
48 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
49 |
## Evaluation
|
50 |
|
51 |
|
|
|
14 |
# Wav2Vec2Bert Audio frame classifier for prosodic unit detection
|
15 |
|
16 |
This model predicts prosodic units on speech.
|
17 |
+
For each 20ms frame the model predicts 1 or 0, indicating whether there is a prosodic unit in
|
18 |
this frame or not.
|
19 |
|
20 |
|
|
|
40 |
|
41 |
## Uses
|
42 |
|
43 |
+
### Simple use (short files)
|
44 |
|
45 |
+
For shorter audios that fit on your GPU the classifier can be used directly.
|
46 |
+
```python
|
47 |
+
import numpy as np
|
48 |
|
49 |
+
from datasets import Audio, Dataset
|
50 |
+
from transformers import AutoFeatureExtractor, Wav2Vec2BertForAudioFrameClassification
|
51 |
+
import torch
|
52 |
+
import numpy as np
|
53 |
+
|
54 |
+
if torch.cuda.is_available():
|
55 |
+
device = torch.device("cuda")
|
56 |
+
else:
|
57 |
+
device = torch.device("cpu")
|
58 |
+
|
59 |
+
model_name = "5roop/Wav2Vec2BertProsodicUnitsFrameClassifier"
|
60 |
+
feature_extractor = AutoFeatureExtractor.from_pretrained(model_name)
|
61 |
+
model = Wav2Vec2BertForAudioFrameClassification.from_pretrained(model_name).to(device)
|
62 |
+
f = "data/Rog-Art-N-G6007-P600702_181.070_211.070.wav"
|
63 |
+
|
64 |
+
|
65 |
+
def frames_to_intervals(frames: list) -> list[tuple]:
|
66 |
+
from itertools import pairwise
|
67 |
+
import pandas as pd
|
68 |
+
|
69 |
+
results = []
|
70 |
+
ndf = pd.DataFrame(
|
71 |
+
data={
|
72 |
+
"time_s": [0.020 * i for i in range(len(frames))],
|
73 |
+
"frames": frames,
|
74 |
+
}
|
75 |
+
)
|
76 |
+
ndf = ndf.dropna()
|
77 |
+
indices_of_change = ndf.frames.diff()[ndf.frames.diff() != 0].index.values
|
78 |
+
for si, ei in pairwise(indices_of_change):
|
79 |
+
if ndf.loc[si : ei - 1, "frames"].mode()[0] == 0:
|
80 |
+
pass
|
81 |
+
else:
|
82 |
+
results.append(
|
83 |
+
(round(ndf.loc[si, "time_s"], 3), round(ndf.loc[ei - 1, "time_s"], 3))
|
84 |
+
)
|
85 |
+
return results
|
86 |
+
|
87 |
+
|
88 |
+
def evaluator(chunks):
|
89 |
+
sampling_rate = chunks["audio"][0]["sampling_rate"]
|
90 |
+
with torch.no_grad():
|
91 |
+
inputs = feature_extractor(
|
92 |
+
[i["array"] for i in chunks["audio"]],
|
93 |
+
return_tensors="pt",
|
94 |
+
sampling_rate=sampling_rate,
|
95 |
+
).to(device)
|
96 |
+
logits = model(**inputs).logits
|
97 |
+
y_pred_raw = np.array(logits.cpu())
|
98 |
+
y_pred = y_pred_raw.argmax(axis=-1)
|
99 |
+
prosodic_units = [frames_to_intervals(i) for i in y_pred]
|
100 |
+
return {
|
101 |
+
"y_pred": y_pred,
|
102 |
+
"y_pred_logits": y_pred_raw,
|
103 |
+
"prosodic_units": prosodic_units,
|
104 |
+
}
|
105 |
+
|
106 |
+
|
107 |
+
ds = Dataset.from_dict({"audio": [f, f]}).cast_column("audio", Audio(16000, mono=True))
|
108 |
+
ds = ds.map(evaluator, batched=True, batch_size=2)
|
109 |
+
print(ds["y_pred"][0])
|
110 |
+
# Outputs: [0, 0, 1, 1, 1, 1, 1, ...]
|
111 |
+
print(ds["y_pred_logits"][0])
|
112 |
+
# Outputs:
|
113 |
+
# [[ 0.89419061, -0.77746612],
|
114 |
+
# [ 0.44213724, -0.34862748],
|
115 |
+
# [-0.08605709, 0.13012762],
|
116 |
+
# ....
|
117 |
+
print(ds["prosodic_units"][0])
|
118 |
+
# Outputs: [[0.04, 2.4], [3.52, 6.6], ....
|
119 |
+
```
|
120 |
+
|
121 |
+
|
122 |
+
### Inference on longer files
|
123 |
+
If the file is too big for straight-forward inference, some chunking needs to be performed in order to process it.
|
124 |
+
We know that for starts and ends of chunks the probability of false negatives increases, so it is best to process the file
|
125 |
+
with some overlap between chunks or split it on silence. We illustrate the former approach here:
|
126 |
+
```python
|
127 |
+
import numpy as np
|
128 |
+
|
129 |
+
from datasets import Audio, Dataset
|
130 |
+
from transformers import AutoFeatureExtractor, Wav2Vec2BertForAudioFrameClassification
|
131 |
+
import torch
|
132 |
+
import numpy as np
|
133 |
+
|
134 |
+
if torch.cuda.is_available():
|
135 |
+
device = torch.device("cuda")
|
136 |
+
else:
|
137 |
+
device = torch.device("cpu")
|
138 |
+
|
139 |
+
model_name = "5roop/Wav2Vec2BertProsodicUnitsFrameClassifier"
|
140 |
+
feature_extractor = AutoFeatureExtractor.from_pretrained(model_name)
|
141 |
+
model = Wav2Vec2BertForAudioFrameClassification.from_pretrained(model_name).to(device)
|
142 |
+
f = "ROG/ROG-Art/WAV/Rog-Art-N-G5025-P600022.wav"
|
143 |
+
|
144 |
+
OVERLAP_S = 10
|
145 |
+
CHUNK_LENGTH_S = 30
|
146 |
+
SAMPLING_RATE = 16_000
|
147 |
+
OVERLAP_SAMPLES = OVERLAP_S * SAMPLING_RATE
|
148 |
+
CHUNK_LENGTH_SAMPLES = CHUNK_LENGTH_S * SAMPLING_RATE
|
149 |
+
|
150 |
+
|
151 |
+
def frames_to_intervals(frames: list) -> list[tuple]:
|
152 |
+
from itertools import pairwise
|
153 |
+
import pandas as pd
|
154 |
+
|
155 |
+
results = []
|
156 |
+
ndf = pd.DataFrame(
|
157 |
+
data={
|
158 |
+
"time_s": [0.020 * i for i in range(len(frames))],
|
159 |
+
"frames": frames,
|
160 |
+
}
|
161 |
+
)
|
162 |
+
ndf = ndf.dropna()
|
163 |
+
indices_of_change = ndf.frames.diff()[ndf.frames.diff() != 0].index.values
|
164 |
+
for si, ei in pairwise(indices_of_change):
|
165 |
+
if ndf.loc[si : ei - 1, "frames"].mode()[0] == 0:
|
166 |
+
pass
|
167 |
+
else:
|
168 |
+
results.append(
|
169 |
+
(round(ndf.loc[si, "time_s"], 3), round(ndf.loc[ei - 1, "time_s"], 3))
|
170 |
+
)
|
171 |
+
return results
|
172 |
+
|
173 |
+
|
174 |
+
def merge_events(events: list[list[float]], centroids):
|
175 |
+
flattened_events = []
|
176 |
+
flattened_centroids = []
|
177 |
+
for batch_idx, batch in enumerate(events):
|
178 |
+
for event in batch:
|
179 |
+
flattened_events.append(event)
|
180 |
+
flattened_centroids.append(centroids[batch_idx])
|
181 |
+
flattened_events.sort(key=lambda x: x[0])
|
182 |
+
|
183 |
+
# Merged list to store final intervals
|
184 |
+
merged = []
|
185 |
+
|
186 |
+
for event, centroid in zip(flattened_events, centroids):
|
187 |
+
if not merged:
|
188 |
+
# If merged is empty, simply add the first event
|
189 |
+
merged.append((event, centroid))
|
190 |
+
else:
|
191 |
+
last_event, last_centroid = merged[-1]
|
192 |
+
# Check for overlap
|
193 |
+
if (last_event[0] < event[1]) and (last_event[1] > event[0]):
|
194 |
+
# Calculate the midpoint of the intervals
|
195 |
+
last_event_midpoint = (last_event[0] + last_event[1]) / 2
|
196 |
+
current_event_midpoint = (event[0] + event[1]) / 2
|
197 |
+
|
198 |
+
# Choose the event whose centroid is closer to its midpoint
|
199 |
+
if abs(last_centroid - last_event_midpoint) <= abs(
|
200 |
+
centroid - current_event_midpoint
|
201 |
+
):
|
202 |
+
continue
|
203 |
+
else:
|
204 |
+
merged[-1] = (event, centroid)
|
205 |
+
else:
|
206 |
+
merged.append((event, centroid))
|
207 |
+
|
208 |
+
final_intervals = [event for event, _ in merged]
|
209 |
+
return final_intervals
|
210 |
+
|
211 |
+
|
212 |
+
def evaluator(chunks):
|
213 |
+
with torch.no_grad():
|
214 |
+
samples = []
|
215 |
+
for array, start, end in zip(chunks["audio"], chunks["start"], chunks["end"]):
|
216 |
+
samples.append(array["array"][start:end])
|
217 |
+
inputs = feature_extractor(
|
218 |
+
samples,
|
219 |
+
return_tensors="pt",
|
220 |
+
sampling_rate=SAMPLING_RATE,
|
221 |
+
).to(device)
|
222 |
+
logits = model(**inputs).logits
|
223 |
+
y_pred_raw = np.array(logits.cpu())
|
224 |
+
y_pred = y_pred_raw.argmax(axis=-1)
|
225 |
+
prosodic_units = [
|
226 |
+
np.array(frames_to_intervals(i)) + start / SAMPLING_RATE
|
227 |
+
for i, start in zip(y_pred, chunks["start"])
|
228 |
+
]
|
229 |
+
return {
|
230 |
+
"y_pred": y_pred,
|
231 |
+
"y_pred_logits": y_pred_raw,
|
232 |
+
"prosodic_units": prosodic_units,
|
233 |
+
}
|
234 |
+
|
235 |
+
|
236 |
+
audio_duration_samples = (
|
237 |
+
Audio(SAMPLING_RATE, mono=True)
|
238 |
+
.decode_example({"path": f, "bytes": None})["array"]
|
239 |
+
.shape[0]
|
240 |
+
)
|
241 |
+
chunk_starts = np.arange(
|
242 |
+
0, audio_duration_samples, CHUNK_LENGTH_SAMPLES - OVERLAP_SAMPLES
|
243 |
+
)
|
244 |
+
chunk_ends = chunk_starts + CHUNK_LENGTH_SAMPLES
|
245 |
+
|
246 |
+
ds = Dataset.from_dict(
|
247 |
+
{
|
248 |
+
"audio": [f for i in chunk_starts],
|
249 |
+
"start": chunk_starts,
|
250 |
+
"end": chunk_ends,
|
251 |
+
"chunk_centroid_s": (chunk_starts + chunk_ends) / 2 / SAMPLING_RATE,
|
252 |
+
}
|
253 |
+
).cast_column("audio", Audio(SAMPLING_RATE, mono=True))
|
254 |
+
|
255 |
+
ds = ds.map(evaluator, batched=True, batch_size=10)
|
256 |
+
|
257 |
+
|
258 |
+
final_intervals = merge_events(ds["prosodic_units"], ds["chunk_centroid_s"])
|
259 |
+
print(final_intervals)
|
260 |
+
# Outputs: [[3.14, 4.96], [5.6, 8.4], [8.62, 9.32], [10.12, 10.7], [11.72, 13.1],....
|
261 |
+
```
|
262 |
## Bias, Risks, and Limitations
|
263 |
|
264 |
## Training Details
|
265 |
|
266 |
+
|hyperparameter|value|
|
267 |
+
|---|---|
|
268 |
+
|learning rate|3e-5|
|
269 |
+
|batch size|1|
|
270 |
+
|gradient accumulation steps|16|
|
271 |
+
|num train epochs|20|
|
272 |
+
|weight decay|0.01|
|
273 |
+
|
274 |
## Evaluation
|
275 |
|
276 |
|