Create app.py
Browse files
app.py
ADDED
|
@@ -0,0 +1,235 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import torch
|
| 2 |
+
import gradio as gr
|
| 3 |
+
from faster_whisper import WhisperModel
|
| 4 |
+
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, pipeline
|
| 5 |
+
from pydub import AudioSegment
|
| 6 |
+
import yt_dlp as youtube_dl
|
| 7 |
+
import tempfile
|
| 8 |
+
from transformers.pipelines.audio_utils import ffmpeg_read
|
| 9 |
+
from gradio.components import Audio, Dropdown, Radio, Textbox
|
| 10 |
+
import os
|
| 11 |
+
import numpy as np
|
| 12 |
+
import soundfile as sf
|
| 13 |
+
os.environ["TOKENIZERS_PARALLELISM"] = "false"
|
| 14 |
+
|
| 15 |
+
|
| 16 |
+
# Paramètres
|
| 17 |
+
FILE_LIMIT_MB = 1000
|
| 18 |
+
YT_LENGTH_LIMIT_S = 3600 # Limite de 1 heure pour les vidéos YouTube
|
| 19 |
+
|
| 20 |
+
# Charger les codes de langue
|
| 21 |
+
from flores200_codes import flores_codes
|
| 22 |
+
|
| 23 |
+
# Fonction pour déterminer le device
|
| 24 |
+
def set_device():
|
| 25 |
+
return torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
| 26 |
+
|
| 27 |
+
device = set_device()
|
| 28 |
+
|
| 29 |
+
|
| 30 |
+
# Charger les modèles une seule fois
|
| 31 |
+
model_dict = {}
|
| 32 |
+
def load_models():
|
| 33 |
+
global model_dict
|
| 34 |
+
if not model_dict:
|
| 35 |
+
model_name_dict = {
|
| 36 |
+
#'nllb-distilled-1.3B': 'facebook/nllb-200-distilled-1.3B',
|
| 37 |
+
'nllb-distilled-600M': 'facebook/nllb-200-distilled-600M',
|
| 38 |
+
#'nllb-1.3B': 'facebook/nllb-200-1.3B',
|
| 39 |
+
#'nllb-distilled-1.3B': 'facebook/nllb-200-distilled-1.3B',
|
| 40 |
+
#'nllb-3.3B': 'facebook/nllb-200-3.3B',
|
| 41 |
+
# 'nllb-distilled-600M': 'facebook/nllb-200-distilled-600M',
|
| 42 |
+
}
|
| 43 |
+
for call_name, real_name in model_name_dict.items():
|
| 44 |
+
model = AutoModelForSeq2SeqLM.from_pretrained(real_name)
|
| 45 |
+
tokenizer = AutoTokenizer.from_pretrained(real_name)
|
| 46 |
+
model_dict[call_name+'_model'] = model
|
| 47 |
+
model_dict[call_name+'_tokenizer'] = tokenizer
|
| 48 |
+
|
| 49 |
+
load_models()
|
| 50 |
+
|
| 51 |
+
model_size = "large-v2"
|
| 52 |
+
model = WhisperModel(model_size)
|
| 53 |
+
|
| 54 |
+
|
| 55 |
+
# Fonction pour la transcription
|
| 56 |
+
def transcribe_audio(audio_file):
|
| 57 |
+
# model_size = "large-v2"
|
| 58 |
+
# model = WhisperModel(model_size)
|
| 59 |
+
# model = WhisperModel(model_size, device=device, compute_type="int8")
|
| 60 |
+
global model
|
| 61 |
+
segments, _ = model.transcribe(audio_file, beam_size=1)
|
| 62 |
+
transcriptions = [("[%.2fs -> %.2fs]" % (seg.start, seg.end), seg.text) for seg in segments]
|
| 63 |
+
return transcriptions
|
| 64 |
+
|
| 65 |
+
|
| 66 |
+
# Fonction pour la traduction
|
| 67 |
+
def traduction(text, source_lang, target_lang):
|
| 68 |
+
# Vérifier si les codes de langue sont dans flores_codes
|
| 69 |
+
if source_lang not in flores_codes or target_lang not in flores_codes:
|
| 70 |
+
print(f"Code de langue non trouvé : {source_lang} ou {target_lang}")
|
| 71 |
+
return ""
|
| 72 |
+
|
| 73 |
+
src_code = flores_codes[source_lang]
|
| 74 |
+
tgt_code = flores_codes[target_lang]
|
| 75 |
+
|
| 76 |
+
model_name = "nllb-distilled-600M"
|
| 77 |
+
model = model_dict[model_name + "_model"]
|
| 78 |
+
tokenizer = model_dict[model_name + "_tokenizer"]
|
| 79 |
+
translator = pipeline("translation", model=model, tokenizer=tokenizer)
|
| 80 |
+
|
| 81 |
+
return translator(text, src_lang=src_code, tgt_lang=tgt_code)[0]["translation_text"]
|
| 82 |
+
|
| 83 |
+
|
| 84 |
+
# Fonction principale
|
| 85 |
+
def full_transcription_and_translation(audio_input, source_lang, target_lang):
|
| 86 |
+
# Si audio_input est une URL
|
| 87 |
+
if isinstance(audio_input, str) and audio_input.startswith("http"):
|
| 88 |
+
audio_file = download_yt_audio(audio_input)
|
| 89 |
+
# Si audio_input est un dictionnaire contenant des données audio
|
| 90 |
+
elif isinstance(audio_input, dict) and "array" in audio_input and "sampling_rate" in audio_input:
|
| 91 |
+
audio_array = audio_input["array"]
|
| 92 |
+
sampling_rate = audio_input["sampling_rate"]
|
| 93 |
+
# Écrire le tableau NumPy dans un fichier temporaire WAV
|
| 94 |
+
with tempfile.NamedTemporaryFile(delete=False, suffix='.wav') as f:
|
| 95 |
+
sf.write(f, audio_array, sampling_rate)
|
| 96 |
+
audio_file = f.name
|
| 97 |
+
else:
|
| 98 |
+
# Supposons que c'est un chemin de fichier
|
| 99 |
+
audio_file = audio_input
|
| 100 |
+
|
| 101 |
+
transcriptions = transcribe_audio(audio_file)
|
| 102 |
+
translations = [(timestamp, traduction(text, source_lang, target_lang)) for timestamp, text in transcriptions]
|
| 103 |
+
|
| 104 |
+
# Supprimez le fichier temporaire s'il a été créé
|
| 105 |
+
if isinstance(audio_input, dict):
|
| 106 |
+
os.remove(audio_file)
|
| 107 |
+
|
| 108 |
+
return transcriptions, translations
|
| 109 |
+
|
| 110 |
+
# Téléchargement audio YouTube
|
| 111 |
+
"""def download_yt_audio(yt_url):
|
| 112 |
+
with tempfile.NamedTemporaryFile(suffix='.mp3') as f:
|
| 113 |
+
ydl_opts = {
|
| 114 |
+
'format': 'bestaudio/best',
|
| 115 |
+
'outtmpl': f.name,
|
| 116 |
+
'postprocessors': [{
|
| 117 |
+
'key': 'FFmpegExtractAudio',
|
| 118 |
+
'preferredcodec': 'mp3',
|
| 119 |
+
'preferredquality': '192',
|
| 120 |
+
}],
|
| 121 |
+
}
|
| 122 |
+
with youtube_dl.YoutubeDL(ydl_opts) as ydl:
|
| 123 |
+
ydl.download([yt_url])
|
| 124 |
+
return f.name"""
|
| 125 |
+
|
| 126 |
+
lang_codes = list(flores_codes.keys())
|
| 127 |
+
|
| 128 |
+
# Interface Gradio
|
| 129 |
+
def gradio_interface(audio_file, source_lang, target_lang):
|
| 130 |
+
if audio_file.startswith("http"):
|
| 131 |
+
audio_file = download_yt_audio(audio_file)
|
| 132 |
+
transcriptions, translations = full_transcription_and_translation(audio_file, source_lang, target_lang)
|
| 133 |
+
transcribed_text = '\n'.join([f"{timestamp}: {text}" for timestamp, text in transcriptions])
|
| 134 |
+
translated_text = '\n'.join([f"{timestamp}: {text}" for timestamp, text in translations])
|
| 135 |
+
return transcribed_text, translated_text
|
| 136 |
+
|
| 137 |
+
|
| 138 |
+
def _return_yt_html_embed(yt_url):
|
| 139 |
+
video_id = yt_url.split("?v=")[-1]
|
| 140 |
+
HTML_str = (
|
| 141 |
+
f'<center> <iframe width="500" height="320" src="https://www.youtube.com/embed/{video_id}"> </iframe>'
|
| 142 |
+
" </center>"
|
| 143 |
+
)
|
| 144 |
+
return HTML_str
|
| 145 |
+
|
| 146 |
+
def download_yt_audio(yt_url, filename):
|
| 147 |
+
info_loader = youtube_dl.YoutubeDL()
|
| 148 |
+
|
| 149 |
+
try:
|
| 150 |
+
info = info_loader.extract_info(yt_url, download=False)
|
| 151 |
+
except youtube_dl.utils.DownloadError as err:
|
| 152 |
+
raise gr.Error(str(err))
|
| 153 |
+
|
| 154 |
+
file_length = info["duration_string"]
|
| 155 |
+
file_h_m_s = file_length.split(":")
|
| 156 |
+
file_h_m_s = [int(sub_length) for sub_length in file_h_m_s]
|
| 157 |
+
|
| 158 |
+
if len(file_h_m_s) == 1:
|
| 159 |
+
file_h_m_s.insert(0, 0)
|
| 160 |
+
if len(file_h_m_s) == 2:
|
| 161 |
+
file_h_m_s.insert(0, 0)
|
| 162 |
+
file_length_s = file_h_m_s[0] * 3600 + file_h_m_s[1] * 60 + file_h_m_s[2]
|
| 163 |
+
|
| 164 |
+
if file_length_s > YT_LENGTH_LIMIT_S:
|
| 165 |
+
yt_length_limit_hms = time.strftime("%HH:%MM:%SS", time.gmtime(YT_LENGTH_LIMIT_S))
|
| 166 |
+
file_length_hms = time.strftime("%HH:%MM:%SS", time.gmtime(file_length_s))
|
| 167 |
+
raise gr.Error(f"Maximum YouTube length is {yt_length_limit_hms}, got {file_length_hms} YouTube video.")
|
| 168 |
+
|
| 169 |
+
ydl_opts = {"outtmpl": filename, "format": "worstvideo[ext=mp4]+bestaudio[ext=m4a]/best[ext=mp4]/best"}
|
| 170 |
+
|
| 171 |
+
with youtube_dl.YoutubeDL(ydl_opts) as ydl:
|
| 172 |
+
try:
|
| 173 |
+
ydl.download([yt_url])
|
| 174 |
+
except youtube_dl.utils.ExtractorError as err:
|
| 175 |
+
raise gr.Error(str(err))
|
| 176 |
+
|
| 177 |
+
|
| 178 |
+
def yt_transcribe(yt_url, task, max_filesize=75.0):
|
| 179 |
+
html_embed_str = _return_yt_html_embed(yt_url)
|
| 180 |
+
global model # S'assurer que le modèle est accessible
|
| 181 |
+
|
| 182 |
+
with tempfile.TemporaryDirectory() as tmpdirname:
|
| 183 |
+
filepath = os.path.join(tmpdirname, "video.mp4")
|
| 184 |
+
download_yt_audio(yt_url, filepath)
|
| 185 |
+
with open(filepath, "rb") as f:
|
| 186 |
+
inputs = f.read()
|
| 187 |
+
|
| 188 |
+
inputs = ffmpeg_read(inputs, model.feature_extractor.sampling_rate)
|
| 189 |
+
inputs = {"array": inputs, "sampling_rate": model.feature_extractor.sampling_rate}
|
| 190 |
+
|
| 191 |
+
transcriptions, translations = full_transcription_and_translation(inputs, source_lang, target_lang)
|
| 192 |
+
transcribed_text = '\n'.join([f"{timestamp}: {text}" for timestamp, text in transcriptions])
|
| 193 |
+
translated_text = '\n'.join([f"{timestamp}: {text}" for timestamp, text in translations])
|
| 194 |
+
return html_embed_str, transcribed_text, translated_text
|
| 195 |
+
|
| 196 |
+
|
| 197 |
+
# Interfaces
|
| 198 |
+
demo = gr.Blocks()
|
| 199 |
+
|
| 200 |
+
with demo:
|
| 201 |
+
with gr.Tab("Microphone"):
|
| 202 |
+
gr.Interface(
|
| 203 |
+
fn=gradio_interface,
|
| 204 |
+
inputs=[
|
| 205 |
+
gr.Audio(sources=["microphone"], type="filepath"),
|
| 206 |
+
gr.Dropdown(lang_codes, value='French', label='Source Language'),
|
| 207 |
+
gr.Dropdown(lang_codes, value='English', label='Target Language')],
|
| 208 |
+
outputs=[gr.Textbox(label="Transcribed Text"), gr.Textbox(label="Translated Text")]
|
| 209 |
+
)
|
| 210 |
+
|
| 211 |
+
with gr.Tab("Audio file"):
|
| 212 |
+
gr.Interface(
|
| 213 |
+
fn=gradio_interface,
|
| 214 |
+
inputs=[
|
| 215 |
+
gr.Audio(type="filepath", label="Audio file"),
|
| 216 |
+
gr.Dropdown(lang_codes, value='French', label='Source Language'),
|
| 217 |
+
gr.Dropdown(lang_codes, value='English', label='Target Language')],
|
| 218 |
+
outputs=[gr.Textbox(label="Transcribed Text"), gr.Textbox(label="Translated Text")]
|
| 219 |
+
)
|
| 220 |
+
|
| 221 |
+
with gr.Tab("YouTube"):
|
| 222 |
+
gr.Interface(
|
| 223 |
+
fn=yt_transcribe,
|
| 224 |
+
inputs=[
|
| 225 |
+
gr.Textbox(lines=1, placeholder="Paste the URL to a YouTube video here", label="YouTube URL"),
|
| 226 |
+
gr.Dropdown(lang_codes, value='French', label='Source Language'),
|
| 227 |
+
gr.Dropdown(lang_codes, value='English', label='Target Language')
|
| 228 |
+
],
|
| 229 |
+
outputs=["html", gr.Textbox(label="Transcribed Text"), gr.Textbox(label="Translated Text")]
|
| 230 |
+
)
|
| 231 |
+
|
| 232 |
+
#with demo:
|
| 233 |
+
#gr.TabbedInterface([mf_transcribe, file_transcribe, yt_transcribe], ["Microphone", "Audio file", "YouTube"])
|
| 234 |
+
|
| 235 |
+
demo.launch()
|