File size: 15,814 Bytes
9b9c715
 
 
 
 
 
 
 
 
 
 
e154110
318370a
 
 
 
 
e154110
318370a
 
9b9c715
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
08238e0
9b9c715
 
 
 
 
 
 
 
 
 
 
08238e0
 
 
9b9c715
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e154110
9b9c715
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8a2aca3
9b9c715
 
 
 
 
 
 
 
 
 
 
08238e0
9b9c715
 
 
08238e0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9b9c715
 
 
 
08238e0
 
 
9b9c715
 
 
 
 
 
08238e0
9b9c715
 
08238e0
 
9b9c715
08238e0
9b9c715
 
08238e0
 
 
9b9c715
08238e0
 
 
 
 
9b9c715
 
 
 
 
 
318370a
 
9b9c715
 
 
 
 
 
 
 
 
e154110
 
 
9b9c715
 
 
 
 
 
 
 
e154110
9b9c715
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e154110
 
 
9b9c715
 
 
 
e154110
9b9c715
e154110
 
 
9b9c715
318370a
e154110
9b9c715
 
e154110
9b9c715
318370a
e154110
 
 
 
 
 
318370a
e154110
 
 
 
 
318370a
 
e154110
 
 
318370a
 
 
 
 
 
 
 
e154110
318370a
e154110
 
 
 
 
 
 
9b9c715
 
e154110
9b9c715
 
318370a
e154110
 
 
 
9b9c715
318370a
 
9b9c715
318370a
e154110
318370a
 
 
 
 
 
 
 
e154110
 
9b9c715
 
318370a
9b9c715
 
 
e154110
 
 
 
 
9b9c715
 
 
 
e154110
 
 
9b9c715
 
 
 
 
318370a
e154110
 
 
 
 
 
 
 
 
 
9b9c715
e154110
9b9c715
318370a
e154110
 
 
 
9b9c715
318370a
 
9b9c715
e154110
 
 
 
 
 
 
9b9c715
 
318370a
9b9c715
e154110
 
 
 
318370a
9b9c715
 
e154110
9b9c715
 
318370a
e154110
 
 
 
9b9c715
318370a
 
9b9c715
e154110
 
 
 
 
 
 
318370a
 
e154110
 
 
 
9b9c715
 
 
 
 
e154110
 
9b9c715
 
 
318370a
e154110
9b9c715
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
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
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
# 1. engineer_style_foreign_style_vectors.py  # for speed=1 & speed=4
# 2. tts_harvard.py                           # (call inside SHIFT repo - needs StyleTTS msinference.py)
# 3. visualize_tts_pleasantness.py            # figures & audinterface


# Visualises timeseries 11 class for mimic3 human mimic3speed
#
#
# human_770.wav
# mimic3_770.wav
# mimic3_speedup_770.wav
FULL_WAV  = [
    'english_hfullh.wav',
    'english_4x_hfullh.wav',
    'human_hfullh.wav',
    'foreign_hfullh.wav',
    'foreign_4x_hfullh.wav',
                    ]
WIN = 40
HOP = 10
import pandas as pd
import os

import json
import numpy as np
import audonnx
import audb
from pathlib import Path
import transformers
import torch
import audmodel
import audinterface
import matplotlib.pyplot as plt
import audiofile

LABELS = ['arousal', 'dominance', 'valence',
        #    'speech_synthesizer', 'synthetic_singing',
           'Angry',
           'Sad',
           'Happy',
           'Surprise', 
            'Fear', 
            'Disgust', 
            'Contempt', 
            'Neutral'
            ]


config = transformers.Wav2Vec2Config() #finetuning_task='spef2feat_reg')
config.dev = torch.device('cuda:0')
config.dev2 = torch.device('cuda:0')

# def _softmax(x):
#     '''x : (batch, num_class)'''
#     x -= x.max(1, keepdims=True)  # if all -400 then sum(exp(x)) = 0
#     x = np.minimum(-100, x)
#     x = np.exp(x)
#     x /= x.sum(1, keepdims=True)
#     return x

def _softmax(x):
    '''x : (batch, num_class)'''
    x -= x.max(1, keepdims=True)  # if all -400 then sum(exp(x)) = 0
    x = np.maximum(-100, x)
    x = np.exp(x)
    x /= x.sum(1, keepdims=True)
    return x

def _sigmoid(x):
    '''x : (batch, num_class)'''
    return 1 / (1 + np.exp(-x))    


        # --
    # ALL = anger, contempt, disgust, fear, happiness, neutral, no_agreement, other, sadness, surprise
    # plot - unplesant emo 7x emo-categories [anger, contempt, disgust, fear, sadness] for artifical/sped-up/natural
    # plot - pleasant emo [neutral, happiness, surprise]
    # plot - Cubes Natural vs spedup   4x speed
    # plot - synthesizer class audioset


    # https://arxiv.org/pdf/2407.12229
    #  https://arxiv.org/pdf/2312.05187
    # https://arxiv.org/abs/2407.05407
    # https://arxiv.org/pdf/2408.06577
    # https://arxiv.org/pdf/2309.07405

    
# wavs are generated concat and plot time-series?

# for mimic3/mimic3speed/human - concat all 77 and run timeseries with 7s hop 3s
for long_audio in FULL_WAV:
    file_interface = f'timeseries_{long_audio.replace("/", "")}.pkl'
    if not os.path.exists(file_interface):


        print('_______________________________________\nProcessing\n', file_interface, '\n___________')



        # CAT MSP

        from transformers import AutoModelForAudioClassification
        import types
        def _infer(self, x):
            '''x: (batch, audio-samples-16KHz)'''
            x = (x + self.config.mean) / self.config.std  # plus
            x = self.ssl_model(x, attention_mask=None).last_hidden_state
            # pool
            h = self.pool_model.sap_linear(x).tanh()
            w = torch.matmul(h, self.pool_model.attention)
            w = w.softmax(1)
            mu = (x * w).sum(1)
            x = torch.cat(
                [
                    mu,
                    ((x * x * w).sum(1) - mu * mu).clamp(min=1e-7).sqrt()
                ], 1)
            return self.ser_model(x)

        teacher_cat = AutoModelForAudioClassification.from_pretrained(
            '3loi/SER-Odyssey-Baseline-WavLM-Categorical-Attributes',
            trust_remote_code=True  # fun definitions see 3loi/SER-.. repo
        ).to(config.dev2).eval()
        teacher_cat.forward = types.MethodType(_infer, teacher_cat)
        

        # ===================[:]===================== Dawn
        def _prenorm(x, attention_mask=None):
            '''mean/var'''
            if attention_mask is not None:
                N = attention_mask.sum(1, keepdim=True)  # here attn msk is unprocessed just the original input
                x -= x.sum(1, keepdim=True) / N
                var = (x * x).sum(1, keepdim=True) / N

            else:
                x -= x.mean(1, keepdim=True)  # mean is an onnx operator reducemean saves some ops compared to casting integer N to float and the div
                var = (x * x).mean(1, keepdim=True)
            return x / torch.sqrt(var + 1e-7)

        from torch import nn
        from transformers.models.wav2vec2.modeling_wav2vec2 import Wav2Vec2PreTrainedModel, Wav2Vec2Model
        class RegressionHead(nn.Module):
                r"""Classification head."""

                def __init__(self, config):

                    super().__init__()

                    self.dense = nn.Linear(config.hidden_size, config.hidden_size)
                    self.dropout = nn.Dropout(config.final_dropout)
                    self.out_proj = nn.Linear(config.hidden_size, config.num_labels)

                def forward(self, features, **kwargs):

                    x = features
                    x = self.dropout(x)
                    x = self.dense(x)
                    x = torch.tanh(x)
                    x = self.dropout(x)
                    x = self.out_proj(x)

                    return x


        class Dawn(Wav2Vec2PreTrainedModel):
            r"""Speech emotion classifier."""

            def __init__(self, config):

                super().__init__(config)

                self.config = config
                self.wav2vec2 = Wav2Vec2Model(config)
                self.classifier = RegressionHead(config)
                self.init_weights()

            def forward(
                    self,
                    input_values,
                    attention_mask=None,
            ):
                x = _prenorm(input_values, attention_mask=attention_mask)
                outputs = self.wav2vec2(x, attention_mask=attention_mask)
                hidden_states = outputs[0]
                hidden_states = torch.mean(hidden_states, dim=1)
                logits = self.classifier(hidden_states)
                return logits
                # return {'hidden_states': hidden_states,
                #         'logits': logits}
        dawn = Dawn.from_pretrained('audeering/wav2vec2-large-robust-12-ft-emotion-msp-dim').to(config.dev).eval()
        # =======================================







        def process_function(x, sampling_rate, idx):
            '''run audioset ct, adv

                USE onnx teachers
                
                return [synth-speech, synth-singing, 7x, 3x adv] = 11
            '''
            
            # x = x[None , :]  ASaHSuFDCN
            #{0: 'Angry', 1: 'Sad', 2: 'Happy', 3: 'Surprise', 
            #4: 'Fear', 5: 'Disgust', 6: 'Contempt', 7: 'Neutral'}
            #tensor([[0.0015, 0.3651, 0.0593, 0.0315, 0.0600, 0.0125, 0.0319, 0.4382]])
            logits_cat = teacher_cat(torch.from_numpy(x).to(config.dev)).cpu().detach().numpy()
            # USE ALL CATEGORIES
            # --
            # logits_audioset = audioset_model(x, 16000)['logits_sounds']
            # logits_audioset = logits_audioset[:, [7, 35]]  # speech synthesizer synthetic singing
            # --
            logits_adv = dawn(torch.from_numpy(x).to(config.dev)).cpu().detach().numpy() #['logits']
            
            cat = np.concatenate([logits_adv,
                                #   _sigmoid(logits_audioset),
                                    _softmax(logits_cat)],
                                    1)
            print(cat)
            return cat #logits_adv #model(signal, sampling_rate)['logits']    


# ---------------------

        
        interface = audinterface.Feature(
            feature_names=LABELS,
            process_func=process_function,
            # process_func_args={'outputs': 'logits_scene'},
            process_func_applies_sliding_window=False,
            win_dur=WIN,
            hop_dur=HOP,
            sampling_rate=16000,
            resample=True,
            verbose=True,
        )
        df_pred = interface.process_file(long_audio)
        df_pred.to_pickle(file_interface)
    else:
        print(file_interface, 'FOUND')
        # df_pred = pd.read_pickle(file_interface)
        
        
        
# ===============================================================================
# V I S U A L S by loading all 3 pkl - mimic3 - speedup - human pd
#
# ===============================================================================


preds  = {}
SHORTEST_PD = 100000  # segments
for long_audio in FULL_WAV:
    file_interface = f'timeseries_{long_audio.replace("/", "")}.pkl'
    y = pd.read_pickle(file_interface)
    preds[long_audio] = y
    SHORTEST_PD = min(SHORTEST_PD, len(y))

# clean indexes for plot

for k,v in preds.items():
    p = v[:SHORTEST_PD]  # TRuncate extra segments - human is slower than mimic3
    # p = pd.read_pickle(student_file)
    p.reset_index(inplace= True)
    p.drop(columns=['file','start'], inplace=True)
    p.set_index('end', inplace=True)
    # p = p.filter(scene_classes) #['transport', 'indoor', 'outdoor'])
    p.index = p.index.map(mapper = (lambda x: x.total_seconds()))
    preds[k] = p

    # print(p, '\n\n\n\n \n')
    
print(preds.keys(),'p')  




# 2 PLOTS

for lang in ['english', 
             'foreign']:
            

            fig, ax = plt.subplots(nrows=8, ncols=2, figsize=(24,20.7),
                                   gridspec_kw={'hspace': 0, 'wspace': .04})


            

            time_stamp = preds['human_hfullh.wav'].index.to_numpy()
            for j, dim in enumerate(['arousal', 
                                    'dominance', 
                                    'valence']):

                # MIMIC3                      

                ax[j, 0].plot(time_stamp, preds[f'{lang}_hfullh.wav'][dim], 
                            color=(0,104/255,139/255), 
                            label='mean_1', 
                            linewidth=2)
                ax[j, 0].fill_between(time_stamp,

                                0*preds[f'{lang}_hfullh.wav'][dim],
                                preds['human_hfullh.wav'][dim],

                                color=(.2,.2,.2), 
                                alpha=0.244)
                if j == 0:
                    if lang == 'english':
                        desc = 'English'
                    else:
                        desc = 'Non-English'
                    ax[j, 0].legend([f'StyleTTS2 using Mimic-3 {desc}',
                                     f'StyleTTS2 uising EmoDB'], 
                                    prop={'size': 14},
                                    )
                ax[j, 0].set_ylabel(dim.lower(), color=(.4, .4, .4), fontsize=17)
                
                # TICK
                ax[j, 0].set_ylim([1e-7, .9999])
                # ax[j, 0].set_yticks([.25, .5,.75])
                # ax[j, 0].set_yticklabels(['0.25', '.5', '0.75'])
                ax[j, 0].set_xticklabels(['' for _ in ax[j, 0].get_xticklabels()])
                ax[j, 0].set_xlim([time_stamp[0], time_stamp[-1]])


            # MIMIC3   4x speed


                ax[j, 1].plot(time_stamp, preds[f'{lang}_4x_hfullh.wav'][dim], 
                            color=(0,104/255,139/255), 
                            label='mean_1', 
                            linewidth=2)
                ax[j, 1].fill_between(time_stamp,

                                0 * preds[f'{lang}_4x_hfullh.wav'][dim],
                                preds['human_hfullh.wav'][dim],

                                color=(.2,.2,.2),
                                alpha=0.244)
                if j == 0:
                    if lang == 'english':
                        desc = 'English'
                    else:
                        desc = 'Non-English'                 
                    ax[j, 1].legend([f'StyleTTS2 using Mimic-3 {desc} 4x speed',
                                    f'StyleTTS2 using EmoDB'], 
                                    prop={'size': 14}, 
                                    #  loc='lower right'
                                    )


                ax[j, 1].set_xlabel('720 Harvard Sentences')



                # TICK
                ax[j, 1].set_ylim([1e-7, .9999])
                # ax[j, 1].set_yticklabels(['' for _ in ax[j, 1].get_yticklabels()])
                ax[j, 1].set_xticklabels(['' for _ in ax[j, 0].get_xticklabels()])
                ax[j, 1].set_xlim([time_stamp[0], time_stamp[-1]])




                ax[j, 0].grid()
                ax[j, 1].grid()
            # CATEGORIE





            time_stamp = preds['human_hfullh.wav'].index.to_numpy()
            for j, dim in enumerate(['Angry', 
                                    'Sad',
                                    'Happy',
                                    #  'Surprise', 
                                    'Fear',
                                    'Disgust', 
                                    #  'Contempt',
                                    #  'Neutral'
                                    ]):   # ASaHSuFDCN
                j = j + 3  # skip A/D/V suplt                         

                # MIMIC3                      

                ax[j, 0].plot(time_stamp, preds[f'{lang}_hfullh.wav'][dim], 
                            color=(0,104/255,139/255), 
                            label='mean_1', 
                            linewidth=2)
                ax[j, 0].fill_between(time_stamp,

                                0*preds[f'{lang}_hfullh.wav'][dim],
                                preds['human_hfullh.wav'][dim],

                                color=(.2,.2,.2), 
                                alpha=0.244)
                # ax[j, 0].legend(['StyleTTS2 style mimic3',
                #                  'StyleTTS2 style crema-d'], 
                #                  prop={'size': 10}, 
                #                 #  loc='upper left'
                # )


                ax[j, 0].set_ylabel(dim.lower(), color=(.4, .4, .4), fontsize=17)

                # TICKS
                ax[j, 0].set_ylim([1e-7, .9999])
                ax[j, 0].set_xlim([time_stamp[0], time_stamp[-1]])
                ax[j, 0].set_xticklabels(['' for _ in ax[j, 0].get_xticklabels()])
                ax[j, 0].set_xlabel('720 Harvard Sentences', fontsize=17, color=(.2,.2,.2))


            # MIMIC3   4x speed


                ax[j, 1].plot(time_stamp, preds[f'{lang}_4x_hfullh.wav'][dim],
                            color=(0,104/255,139/255), 
                            label='mean_1', 
                            linewidth=2)
                ax[j, 1].fill_between(time_stamp,

                                0*preds[f'{lang}_4x_hfullh.wav'][dim],
                                preds['human_hfullh.wav'][dim],

                                color=(.2,.2,.2), 
                                alpha=0.244)
                # ax[j, 1].legend(['StyleTTS2 style mimic3   4x speed',
                #                  'StyleTTS2 style crema-d'], 
                #                  prop={'size': 10},
                #                 #  loc='upper left'
                # )
                ax[j, 1].set_xlabel('720 Harvard Sentences', fontsize=17, color=(.2,.2,.2))
                ax[j, 1].set_ylim([1e-7, .9999])
                # ax[j, 1].set_yticklabels(['' for _ in ax[j, 1].get_yticklabels()])
                ax[j, 1].set_xticklabels(['' for _ in ax[j, 1].get_xticklabels()])
                ax[j, 1].set_xlim([time_stamp[0], time_stamp[-1]])
                





                ax[j, 0].grid()
                ax[j, 1].grid()



            plt.savefig(f'fig_{lang}_{WIN=}_{HOP=}_fin0.pdf', bbox_inches='tight')
            plt.close()