File size: 7,235 Bytes
da0e3ab
73fd4c0
 
ed64e04
80b43a8
73fd4c0
 
 
fbd6bad
73fd4c0
 
 
 
 
cf7b168
739fd69
ae3f094
 
 
 
 
7b3eb41
73fd4c0
2d49e86
 
 
 
 
9e19d29
fbd6bad
26b0345
fbd6bad
b6ee570
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3364e9c
2275971
79ce0ab
2275971
80b43a8
 
ac401be
992f837
45a4010
10ac59a
233c677
 
 
b6ee570
 
 
 
 
 
79ce0ab
b6ee570
80b43a8
233c677
0cab5bf
 
 
 
233c677
0cab5bf
 
 
 
233c677
0cab5bf
233c677
 
fbd6bad
 
 
902b7eb
58ceda5
233c677
 
 
4eae89a
 
9462754
4eae89a
43edaa1
 
 
233c677
80b43a8
 
6f11b02
233c677
 
 
 
 
 
 
 
5d311f1
 
 
 
b6ee570
 
 
 
 
 
ddb704c
b6ee570
80b43a8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
892ff2a
 
 
3364e9c
 
 
 
 
 
 
6312799
73fd4c0
 
 
3364e9c
 
b0de989
5d311f1
73fd4c0
3364e9c
c3f9f52
 
4bc5468
c3f9f52
73fd4c0
3364e9c
 
 
4bc5468
 
 
 
 
 
 
5d311f1
4bc5468
6783c16
3364e9c
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
import tempfile
import gradio as gr
import subprocess
import os, stat
import uuid
from googletrans import Translator
from TTS.api import TTS
import ffmpeg
from faster_whisper import WhisperModel
from scipy.signal import wiener
import soundfile as sf
from pydub import AudioSegment
import numpy as np
import librosa
from zipfile import ZipFile
import shlex
import cv2
import torch
import torchvision
from tqdm import tqdm
from numba import jit

os.environ["COQUI_TOS_AGREED"] = "1"

ZipFile("ffmpeg.zip").extractall()
st = os.stat('ffmpeg')
os.chmod('ffmpeg', st.st_mode | stat.S_IEXEC)

#Whisper
model_size = "small"
model = WhisperModel(model_size, device="cuda", compute_type="float16")

def check_for_faces(video_path):
    face_cascade = cv2.CascadeClassifier(cv2.data.haarcascades + 'haarcascade_frontalface_default.xml')
    cap = cv2.VideoCapture(video_path)

    while True:
        ret, frame = cap.read()
        if not ret:
            break

        gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
        faces = face_cascade.detectMultiScale(gray, 1.1, 4)

        if len(faces) > 0:
            return True

    return False
    
def process_video(radio, video, target_language):
    if target_language is None:
        return gr.Error("Please select a Target Language for Dubbing.")
        
    run_uuid = uuid.uuid4().hex[:6]
    output_filename = f"{run_uuid}_resized_video.mp4"
    ffmpeg.input(video).output(output_filename, vf='scale=-2:720').run()

    video_path = output_filename
    
    if not os.path.exists(video_path):
        return f"Error: {video_path} does not exist."

    # Move the duration check here
    video_info = ffmpeg.probe(video_path)
    video_duration = float(video_info['streams'][0]['duration'])

    if video_duration > 60:
        os.remove(video_path)  # Delete the resized video
        return gr.Error("Video duration exceeds 1 minute. Please upload a shorter video.")

    ffmpeg.input(video_path).output(f"{run_uuid}_output_audio.wav", acodec='pcm_s24le', ar=48000, map='a').run()

    #y, sr = sf.read(f"{run_uuid}_output_audio.wav")
    #y = y.astype(np.float32)
    #y_denoised = wiener(y)
    #sf.write(f"{run_uuid}_output_audio_denoised.wav", y_denoised, sr)

    #sound = AudioSegment.from_file(f"{run_uuid}_output_audio_denoised.wav", format="wav")
    #sound = sound.apply_gain(0)
    #sound = sound.low_pass_filter(3000).high_pass_filter(100)
    #sound.export(f"{run_uuid}_output_audio_processed.wav", format="wav")

    shell_command = f"ffmpeg -y -i {run_uuid}_output_audio.wav -af lowpass=3000,highpass=100 {run_uuid}_output_audio_final.wav".split(" ")
    subprocess.run([item for item in shell_command], capture_output=False, text=True, check=True)

    segments, info = model.transcribe(f"{run_uuid}_output_audio_final.wav", beam_size=5)
    whisper_text = " ".join(segment.text for segment in segments)
    whisper_language = info.language
    print(whisper_text)
    
    language_mapping = {'English': 'en', 'Spanish': 'es', 'French': 'fr', 'German': 'de', 'Italian': 'it', 'Portuguese': 'pt', 'Polish': 'pl', 'Turkish': 'tr', 'Russian': 'ru', 'Dutch': 'nl', 'Czech': 'cs', 'Arabic': 'ar', 'Chinese (Simplified)': 'zh-cn'}
    target_language_code = language_mapping[target_language]
    translator = Translator()
    try:
        translated_text = translator.translate(whisper_text, src=whisper_language, dest=target_language_code).text
        print(translated_text)
    except AttributeError as e:
        print("Failed to translate text. Likely an issue with token extraction in the Google Translate API.")
        translated_text = "Translation failed due to API issue."

    tts = TTS("tts_models/multilingual/multi-dataset/xtts_v1")
    tts.to('cuda')
    tts.tts_to_file(translated_text, speaker_wav=f"{run_uuid}_output_audio_final.wav", file_path=f"{run_uuid}_output_synth.wav", language=target_language_code)
    
    pad_top = 0
    pad_bottom = 15
    pad_left = 0
    pad_right = 0
    rescaleFactor = 1

    video_path_fix = video_path

    if has_closeup_face:
        has_face = True
    else:
        has_face = check_for_faces(video_path)

    if has_face:
        cmd = f"python Wav2Lip/inference.py --checkpoint_path 'Wav2Lip/checkpoints/wav2lip_gan.pth' --face {shlex.quote(video_path)} --audio '{run_uuid}_output_synth.wav' --pads {pad_top} {pad_bottom} {pad_left} {pad_right} --resize_factor {rescaleFactor} --nosmooth --outfile '{run_uuid}_output_video.mp4'"
        subprocess.run(cmd, shell=True)
    else:
        # Merge audio with the original video without running Wav2Lip
        cmd = f"ffmpeg -i {video_path} -i {run_uuid}_output_synth.wav -c:v copy -c:a aac -strict experimental -map 0:v:0 -map 1:a:0 {run_uuid}_output_video.mp4"
        subprocess.run(cmd, shell=True)

    if not os.path.exists(f"{run_uuid}_output_video.mp4"):
        raise FileNotFoundError(f"Error: {run_uuid}_output_video.mp4 was not generated.")

    output_video_path = f"{run_uuid}_output_video.mp4"

    # Cleanup: Delete all generated files except the final output video
    files_to_delete = [
        f"{run_uuid}_resized_video.mp4",
        f"{run_uuid}_output_audio.wav",
        f"{run_uuid}_output_audio_final.wav",
        f"{run_uuid}_output_synth.wav"
    ]
    for file in files_to_delete:
        try:
            os.remove(file)
        except FileNotFoundError:
            print(f"File {file} not found for deletion.")

    return output_video_path
    
    
def swap(radio):
    if(radio == "Upload"):
        return gr.update(source="upload")
    else:
        return gr.update(source="webcam")
        
video = gr.Video()
radio = gr.Radio(["Upload", "Record"], value="Upload", show_label=False)
iface = gr.Interface(
    fn=process_video,
    inputs=[
        radio,
        video,
        gr.Dropdown(choices=["English", "Spanish", "French", "German", "Italian", "Portuguese", "Polish", "Turkish", "Russian", "Dutch", "Czech", "Arabic", "Chinese (Simplified)"], label="Target Language for Dubbing", value="Spanish")
        checkbox = gr.Checkbox(label="Video has a close-up face", default=False)    
    ],
    outputs=gr.Video(),
    live=False,
    title="AI Video Dubbing",
    description="""This tool was developed by [@artificialguybr](https://twitter.com/artificialguybr) using entirely open-source tools. Special thanks to Hugging Face for the GPU support. Thanks [@yeswondwer](https://twitter.com/@yeswondwerr) for original code.""",
    allow_flagging=False
)
with gr.Blocks() as demo:
    iface.render()
    radio.change(swap, inputs=[radio], outputs=video)
    gr.Markdown("""
    **Note:**
    - Video limit is 1 minute. It will dubbling all people using just one voice.
    - Generation may take up to 5 minutes.
    - The tool uses open-source models for all models. It's a alpha version.
    - Quality can be improved but would require more processing time per video. For scalability and hardware limitations, speed was chosen, not just quality.
    - If you need more than 1 minute, duplicate the Space and change the limit on app.py.
    - If you incorrectly mark the 'Video has a close-up face' checkbox, the dubbing may not work as expected.
    """)
demo.queue(concurrency_count=1, max_size=15)
demo.launch()