File size: 7,648 Bytes
b519f35
 
b978cf1
fcb97b9
 
18d9abd
fcb97b9
 
 
 
 
 
 
 
 
b519f35
fcb97b9
 
 
 
 
 
 
 
 
 
 
 
 
ebd3dc2
7ea9a86
43b7ecc
 
fcb97b9
 
 
 
 
 
 
 
 
 
 
 
 
 
8d72588
9f8665b
fcb97b9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9f8665b
fcb97b9
 
 
 
 
c911c9e
fcb97b9
 
 
 
2d8df4b
fcb97b9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
bd61fdc
fcb97b9
 
4ce50c9
 
fcb97b9
 
a6e3817
7936882
7441b82
a065eb5
fcb97b9
 
 
 
2d8df4b
f78233d
fcb97b9
 
43b7ecc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
88031b7
43b7ecc
 
 
fcb97b9
18d9abd
922104e
 
875e8b8
8d72588
a887a28
 
76d4547
4ce50c9
 
fcb97b9
 
2d8df4b
c911c9e
fcb97b9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4ce50c9
fcb97b9
 
a5292a5
 
fcb97b9
a5292a5
43b7ecc
 
 
 
 
 
 
7441b82
ab8067c
58d5091
ab8067c
58d5091
ab8067c
43b7ecc
fcb97b9
cd01dff
43b7ecc
 
fcb97b9
 
2d8df4b
cd01dff
fcb97b9
2d8df4b
 
 
fcb97b9
 
be84e8f
69eea98
2d8df4b
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
import warnings
warnings.filterwarnings("ignore") # ๊ฒฝ๊ณ  ๋ฌด์‹œ
from moviepy import * #VideoFileClip
#!pip install pyannote.audio
#!pip install moviepy
#!pip install gradio
import librosa
import numpy as np
import os

import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader, TensorDataset
import torch.functional as F
import torch.nn.functional as F

from pyannote.audio import Pipeline

from pyannote.audio import Audio
import torchaudio


from transformers import pipeline

from huggingface_hub import hf_hub_download

import gradio as gr

import shutil

import speech_recognition as sr


# ์˜ค๋””์˜ค ๋ณ€ํ™˜ mp4 --> wav
def extract_audio_from_video(video_file_path, audio_file_path):
    # mp4 ํŒŒ์ผ ๋ถˆ๋Ÿฌ์˜ค๊ธฐ
    video = VideoFileClip(video_file_path)

    # ์˜ค๋””์˜ค๋ฅผ ์ถ”์ถœํ•˜์—ฌ wav ํŒŒ์ผ๋กœ ์ €์žฅ
    video.audio.write_audiofile(audio_file_path, codec='pcm_s16le')

# ์ „์ฒด ์˜ค๋””์˜ค ํŒŒ์ผ ๋ถˆ๋Ÿฌ์˜ค๊ธฐ
def seprate_speaker(audio_file, pipeline):
    audio = Audio()
    waveform, sample_rate = torchaudio.load(audio_file)
    diarization = pipeline(audio_file)

        
    # ํ™”์ž๋ณ„๋กœ ๋ฐœํ™” ๊ตฌ๊ฐ„์„ ์ €์žฅํ•  ๋”•์…”๋„ˆ๋ฆฌ ์ดˆ๊ธฐํ™”
    speaker_segments = {}

    # diarization ๊ฒฐ๊ณผ๋ฅผ ์ˆœํšŒํ•˜๋ฉฐ ๊ฐ ํ™”์ž์˜ ๋ฐœํ™”๋ฅผ ๋”•์…”๋„ˆ๋ฆฌ์— ์ถ”๊ฐ€
    for segment, _, speaker in diarization.itertracks(yield_label=True):
        start_time = segment.start
        end_time = segment.end

        # ํ•ด๋‹น ํ™”์ž๊ฐ€ ์ฒ˜์Œ ๋“ฑ์žฅํ•˜๋ฉด ๋ฆฌ์ŠคํŠธ๋ฅผ ์ดˆ๊ธฐํ™”
        if speaker not in speaker_segments:
            speaker_segments[speaker] = []

        # ๋ฐœํ™” ๊ตฌ๊ฐ„์„ ํ•ด๋‹น ํ™”์ž์˜ ๋ฆฌ์ŠคํŠธ์— ์ถ”๊ฐ€
        segment_waveform = waveform[:, int(start_time * sample_rate):int(end_time * sample_rate)]
        speaker_segments[speaker].append(segment_waveform)


    # ๊ฐ ํ™”์ž๋ณ„๋กœ ๋ชจ๋“  ๋ฐœํ™” ๊ตฌ๊ฐ„์„ ํ•˜๋‚˜์˜ ํŒŒ์ผ๋กœ ์ด์–ด๋ถ™์—ฌ ์ €์žฅ
    for speaker, segments in speaker_segments.items():
        # ํ™”์ž์˜ ๋ชจ๋“  ๋ฐœํ™” ๊ตฌ๊ฐ„์„ ์ด์–ด๋ถ™์ž„
        combined_waveform = torch.cat(segments, dim=1)
        output_path = "/tmp/wav"    # ๊ฒฝ๋กœ
        os.makedirs(output_path, exist_ok=True) # ๊ฒฝ๋กœ๊ฐ€ ์—†์œผ๋ฉด ์ƒ์„ฑ
        output_filename = os.path.join(output_path,f"{speaker}.wav")

        torchaudio.save(output_filename, combined_waveform, sample_rate) #์˜ค๋””์˜ค ํŒŒ์ผ ์ €์žฅ



# ๊ฐ„๋‹จํ•œ DeepVoice ์Šคํƒ€์ผ ๋ชจ๋ธ ์ •์˜
class DeepVoiceModel(nn.Module):
    def __init__(self, input_dim, hidden_dim, num_classes, dropout_rate=0.3, l2_reg=0.01):
        super(DeepVoiceModel, self).__init__()
        self.conv1 = nn.Conv1d(input_dim, hidden_dim, kernel_size=5, padding=2)
        self.bn1 = nn.BatchNorm1d(hidden_dim)
        self.conv2 = nn.Conv1d(hidden_dim, hidden_dim, kernel_size=5, padding=2)
        self.bn2 = nn.BatchNorm1d(hidden_dim)
        self.dropout = nn.Dropout(dropout_rate)
        self.fc = nn.Linear(hidden_dim, num_classes)

    def forward(self, x):
        x = self.bn1(torch.relu(self.conv1(x)))
        x = self.dropout(x)
        x = self.bn2(torch.relu(self.conv2(x)))
        x = self.dropout(x)
        x = torch.mean(x, dim=2)  # Temporal pooling
        x = self.fc(x)
        return x


def extract_mfcc_path(file_path, n_mfcc=13, max_len=100):
    # ์Œ์„ฑ ํŒŒ์ผ
    audio, sample_rate = librosa.load(file_path, sr=None)
    # mfcc ํŠน์„ฑ ์ถ”์ถœ
    mfcc = librosa.feature.mfcc(y=audio, sr=sample_rate, n_mfcc=n_mfcc)

    # ์ผ์ •ํ•œ ๊ธธ์ด๋กœ ๋งž์ถค
    if mfcc.shape[1] < max_len:
        pad_width = max_len - mfcc.shape[1]
        mfcc = np.pad(mfcc, ((0, 0), (0, pad_width)), mode='constant')
    else:
        mfcc = mfcc[:, :max_len]

    return torch.Tensor(mfcc)

# ํด๋”์— ์žˆ๋Š” ๋ฐ์ดํ„ฐ ํ•œ๋ฒˆ์— ์ ‘๊ทผํ•ด์„œ ํ•œ๋ฒˆ์— ์ฒดํฌ
def real_fake_check(list_dir, path, model):
    THRESHOLD = 0.4     #๋”ฅํŽ˜์ดํฌ ๊ธฐ์ค€์„ 0.4๋กœ ์„ค์ •
    r_cnt = 0
    f_cnt = 0
    prob = [] 
    for i in list_dir:      # real / fake ์„ ํƒ
        input_data = extract_mfcc_path(os.path.join(path, i))
        #input_data = torch.tensor(input_data).unsqueeze(0).to('cuda')  # ๋ฐฐ์น˜ ์ฐจ์›์„ ์ถ”๊ฐ€ํ•˜์—ฌ (1, input_dim, sequence_length)๋กœ ๋งž์ถค
        input_data = torch.tensor(input_data).unsqueeze(0).to('cpu') 
        result = model(input_data.float())
        probabilities = F.softmax(result, dim=1)
        #prob[i]='%.2f'%probabilities[0][1].item()
        #prob[i]=round(probabilities[0][1].item(),2)
        prob.append(probabilities[0][1].item())
        predicted_class = 0 if probabilities[0][0] >= THRESHOLD else 1  # ํ™•๋ฅ ๊ฐ’์ด ๊ธฐ์ค€์น˜๋ณด๋‹ค ํฌ๋‹ค๋ฉด real, ์•„๋‹ˆ๋ฉด fake
        if predicted_class == 0:
            r_cnt += 1
        else:
            f_cnt += 1

    return {'real':r_cnt, 'fake':f_cnt, 'prob': prob}


# ์Œ์„ฑ์—์„œ text ์ถ”์ถœ
def convert_wav_to_text(wav_file_path):
    recognizer = sr.Recognizer()

    # WAV ํŒŒ์ผ ์—ด๊ธฐ
    with sr.AudioFile(wav_file_path) as source:
        print("WAV ํŒŒ์ผ์—์„œ ์˜ค๋””์˜ค๋ฅผ ๋กœ๋“œ ์ค‘...")
        audio_data = recognizer.record(source)  # ์ „์ฒด ์˜ค๋””์˜ค๋ฅผ ๋…น์Œ

        try:
            # Google Web Speech API๋กœ ํ…์ŠคํŠธ ๋ณ€ํ™˜

            text = recognizer.recognize_google(audio_data)

        except sr.UnknownValueError:
            text = 'error'
        except sr.RequestError as e:
            text = 'error'
            
    return text


def main(file_name):
    
    if os.path.exists('/tmp/wav'):
        shutil.rmtree('/tmp/wav')

        
    hf_token = os.getenv("HUGGINGFACE_TOKEN")
    pipeline = Pipeline.from_pretrained("pyannote/speaker-diarization-3.1",use_auth_token=hf_token)
    #device = torch.device('cuda:0') if torch.cuda.is_available() else "cpu"#torch.device('cpu')
    device = torch.device('cpu')
    #device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

    video_file = file_name #deepfake #meganfox.mp4'
    #current_path = os.getcwd()
    audio_file = '/tmp/output_audio.wav'  # ์ €์žฅํ•  ์˜ค๋””์˜ค ํŒŒ์ผ์˜ ๊ฒฝ๋กœ, ์ด๋ฆ„ ์ง€์ •

    extract_audio_from_video(video_file, audio_file)

    seprate_speaker(audio_file,pipeline) # ๋ฐœํ™”์ž ๋ถ„๋ฆฌํ•ด์„œ ํŒŒ์ผ๋กœ ๋งŒ๋“ค๊ธฐ

    mel_dim = 13 # Mel-spectrogram ์ฐจ์›
    num_classes = 2  # ๋ถ„๋ฅ˜ํ•  ํด๋ž˜์Šค ์ˆ˜
    input_dim = mel_dim
    hidden_dim = 128
    dropout_rate = 0.2
    l2_reg = 0.01

    # ๋ชจ๋ธ
    model_name = hf_hub_download(repo_id="sssssungk/deepfake_voice", filename="deepvoice_model_girl.pth")
    model = DeepVoiceModel(input_dim, hidden_dim, num_classes, dropout_rate, l2_reg).to(device)
    model.load_state_dict(torch.load(model_name, map_location=torch.device('cpu')))
    model.eval()  # ํ‰๊ฐ€ ๋ชจ๋“œ๋กœ ์„ค์ •

    path = '/tmp/wav'
    file_path = os.listdir(path)

    rf_check = real_fake_check(file_path, path,model)      #fake dataset\

    text_list =[]
    to_text = os.listdir('/tmp/wav')
    for i in to_text:
        text = convert_wav_to_text(os.path.join(path, i))
        text_list.append(text)

    text_list.append(convert_wav_to_text(audio_file))    

    # '์ถ”๊ฐ€(" "ํ˜•์‹์œผ๋กœ ๋‚˜์˜ค๊ฒŒ)
    for i in range(len(text_list)): 
        text_list[i] = text_list[i]+"'"
    
    return rf_check, text_list

def deepvoice_check(video_file):
    results,text = main(video_file)
    return results,text

# Gradio ์ธํ„ฐํŽ˜์ด์Šค ์ƒ์„ฑ
deepfake = gr.Interface(
    fn=deepvoice_check,
    inputs=gr.Video(label="Upload mp4 File"),
    outputs=gr.Textbox(label="DeepFaKeVoice Detection Result"),
    title="DeepFaKeVoice Check",
    description="Upload an mp4 file to check."
)


if __name__ == "__main__":
    deepfake.launch(share=True, debug=True)