File size: 9,922 Bytes
e0f5494 26b4ff9 e0f5494 |
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 |
import os
import gradio as gr
import whisperx
import numpy as np
import moviepy.editor as mp
from moviepy.audio.AudioClip import AudioArrayClip
from pytube import YouTube
import deepl
import torch
import pyrubberband as pyrb
import soundfile as sf
import librosa
from TTS.api import TTS
os.environ["COQUI_TOS_AGREED"] = "1"
HF_TOKEN = os.environ["HF_TOKEN"]
DEEPL_TOKEN = os.environ["DEEPL_TOKEN"]
# Download video from Youtube
def download_youtube_video(url):
yt = YouTube(url)
stream = yt.streams.filter(file_extension='mp4').first()
output_path = stream.download()
return output_path
# Extract audio from video
def extract_audio(video_path):
clip = mp.VideoFileClip(video_path)
audio_path = os.path.splitext(video_path)[0] + ".wav"
clip.audio.write_audiofile(audio_path)
return audio_path
# Perform speech diarization
def speech_diarization(audio_path, hf_token):
device = "cuda"
batch_size = 16
compute_type = "float16"
model = whisperx.load_model("large-v2", device, compute_type=compute_type)
# 1. Transcribe audio
audio = whisperx.load_audio(audio_path)
result = model.transcribe(audio, batch_size=batch_size)
# delete model if low on GPU resources
import gc; gc.collect(); torch.cuda.empty_cache(); del model
# 2. Align whisper output
model_a, metadata = whisperx.load_align_model(language_code=result["language"], device=device)
result = whisperx.align(result["segments"], model_a, metadata, audio, device, return_char_alignments=False)
# delete model if low on GPU resources
import gc; gc.collect(); torch.cuda.empty_cache(); del model_a
# 3. Assign speaker labels
diarize_model = whisperx.DiarizationPipeline(model_name='pyannote/[email protected]', use_auth_token=hf_token, device=device)
# add min/max number of speakers if known
diarize_segments = diarize_model(audio)
# diarize_model(audio, min_speakers=min_speakers, max_speakers=max_speakers)
result = whisperx.assign_word_speakers(diarize_segments, result)
print(f'\n[Original transcript]:\n{result["segments"]}\n')
return result["segments"]
# Create per speaker voice clips for tts voice cloning
def speaker_voice_clips(transcription, audio_path):
# Create 3 uninterrupted per speaker timecodes
snippets_timecodes = {}
for segment in transcription:
speaker = segment['speaker']
if speaker not in snippets_timecodes:
snippets_timecodes[speaker] = []
if len(snippets_timecodes[speaker]) < 3:
snippet = {
'start': segment['start'],
'end': segment['end']
}
snippets_timecodes[speaker].append(snippet)
# Cut voice clips and stitch them together
original_audio = mp.AudioFileClip(audio_path)
audio_file_directory = os.path.dirname(audio_path)
voice_clips = {}
for speaker, speaker_snippets in snippets_timecodes.items():
subclips = []
for snippet in speaker_snippets:
start, end = snippet['start'], snippet['end']
subclip = original_audio.subclip(start, end)
subclips.append(subclip)
concatenated_clip = mp.concatenate_audioclips(subclips)
output_filename = os.path.join(audio_file_directory, f"{speaker}_voice_clips.wav")
concatenated_clip.write_audiofile(output_filename)
voice_clips[speaker] = output_filename
return voice_clips
# Perform text translation
def translate_transcript(transcript, target_language, deepl_token):
translator = deepl.Translator(deepl_token)
translated_transcript = []
for segment in transcript:
text_to_translate = segment['text']
translated_text = translator.translate_text(text_to_translate, target_lang=target_language)
translated_segment = {
'start': segment['start'],
'end': segment['end'],
'text': translated_text.text,
'speaker': segment['speaker']
}
translated_transcript.append(translated_segment)
print(f'\n[Translated transcript]:\n{translated_transcript}\n')
return translated_transcript
# Adjust voice pace
def adjust_voice_pace(sound_array, sample_rate, target_duration):
duration = len(sound_array) / sample_rate
tempo_change = duration / target_duration
sound_array_stretched = pyrb.time_stretch(sound_array, sample_rate, tempo_change)
return sound_array_stretched
# Perform voice cloning
def voice_cloning_translation(translated_transcription, speakers_voice_clips, target_language, speaker_model, audio_path):
device = "cuda"
vits_language_map = {
'en':'eng',
'ru':'rus',
'uk':'ukr',
'pl':'pol'
}
# Select model
selected_model = None
if 'vits' in speaker_model.lower() or target_language is 'uk':
selected_model = f'tts_models/{vits_language_map[target_language]}/fairseq/vits'
else:
selected_model = 'tts_models/multilingual/multi-dataset/xtts_v2'
print(selected_model)
tts = None
final_audio_track = None
try:
# TODO uncomment when https://github.com/coqui-ai/TTS/issues/3224 is resolved
# tts = TTS(selected_model).to(device)
# Generate and concatenate voice clips per speaker
last_end_time = 0
clips = []
# Generate sentences
for speech_item in translated_transcription:
speech_item_duration = speech_item['end'] - speech_item['start']
# Silence
gap_duration = speech_item['start'] - last_end_time
if gap_duration > 0:
silent_audio = np.zeros((int(44100 * gap_duration), 2))
silent_clip = AudioArrayClip(silent_audio, fps=44100)
clips.append(silent_clip)
print(f"\nAdded silence: Start={last_end_time}, Duration={gap_duration}")
# Generate speech
print(f"[{speech_item['speaker']}]")
tts = TTS(selected_model).to(device)
audio = tts.tts_with_vc(text=speech_item['text'], speaker_wav=speakers_voice_clips[speech_item['speaker']], language=target_language)
sample_rate = tts.voice_converter.vc_config.audio.output_sample_rate
# Adjust pace to fit the speech timeframe if translated audio is longer than phrase
audio_duration = len(audio) / sample_rate
if speech_item_duration < audio_duration:
audio = adjust_voice_pace(audio, sample_rate, speech_item_duration)
# Resample to higher rate
new_sample_rate = 44100
audio = librosa.resample(audio, orig_sr=sample_rate, target_sr=new_sample_rate)
# Transform to AudioArrayClip object
audio = np.expand_dims(audio, axis=1)
audio_stereo = np.repeat(audio, 2, axis=1)
audio_clip = AudioArrayClip(audio_stereo, fps=44100)
# Cut out possible glitch from AudioArrayClip end
audio_clip = audio_clip.subclip(0, audio_clip.duration - 0.2)
clips.append(audio_clip)
print(f"Added speech: Start={speech_item['start']}, Final duration={audio_clip.duration}, Original duration={speech_item_duration}")
last_end_time = speech_item['start'] + audio_clip.duration
del tts; import gc; gc.collect(); torch.cuda.empty_cache()
# Merge sentences
final_audio_track = mp.concatenate_audioclips(clips)
audio_files_directory = os.path.dirname(audio_path)
final_audio_track.write_audiofile(os.path.join(audio_files_directory, "translated_voice_track.wav"), fps=44100)
except Exception as e:
if tts is not None:
import gc; gc.collect(); torch.cuda.empty_cache(); del tts
raise e
return final_audio_track
def dub_video(video_path, translated_audio_track, target_language):
video = mp.VideoFileClip(video_path)
video = video.subclip(0, translated_audio_track.duration)
original_audio = video.audio.volumex(0.2)
dubbed_audio = mp.CompositeAudioClip([original_audio, translated_audio_track.set_start(0)])
video_with_dubbing = video.set_audio(dubbed_audio)
video_with_dubbing_path = os.path.splitext(video_path)[0] + "_" + target_language + ".mp4"
video_with_dubbing.write_videofile(video_with_dubbing_path)
return video_with_dubbing_path
# Perform video translation
def video_translation(video_path, target_language, speaker_model, hf_token, deepl_token):
original_audio_path = extract_audio(video_path)
transcription = speech_diarization(original_audio_path, hf_token)
translated_transcription = translate_transcript(transcription, target_language, deepl_token)
speakers_voice_clips = speaker_voice_clips(transcription, original_audio_path)
translated_audio_track = voice_cloning_translation(translated_transcription, speakers_voice_clips, target_language, speaker_model, original_audio_path)
video_with_dubbing = dub_video(video_path, translated_audio_track, target_language)
return video_with_dubbing
def translate_video(_, video_path, __, youtube_link, ___, target_language, speaker_model):
try:
if not video_path and not youtube_link:
gr.Warning("You should either upload video or input a YouTube link")
return None
if youtube_link:
video_path = download_youtube_video(youtube_link)
dubbed_video = video_translation(video_path, target_language, speaker_model, HF_TOKEN, DEEPL_TOKEN)
except Exception as e:
print(f"An error occurred: {e}")
return gr.components.Video(dubbed_video)
inputs = [
gr.Markdown("## Currently supported languages are: English, Polish, Ukrainian and Russian"),
gr.Video(label="Upload a video file"),
gr.Markdown("**OR**"),
gr.Textbox(label="Paste YouTube link"),
gr.Markdown("---"),
gr.Dropdown(["en", "pl", "uk", "ru"], value="pl", label="Select translation target language"),
gr.Dropdown(["(Recommended) XTTS_V2", "VITs (will be default for Ukrainian)"], value="(Recommended) XTTS_V2", label="Select text-to-speech generation model")
]
outputs = gr.Video(label="Translated video")
gr.Interface(fn=translate_video,
inputs=inputs,
outputs=outputs,
title="🌐AI Video Translation",
theme=gr.themes.Base()
).launch(show_error=True, debug=True) |