Update app.py
Browse files
app.py
CHANGED
|
@@ -20,7 +20,7 @@ import librosa
|
|
| 20 |
from huggingface_hub import snapshot_download
|
| 21 |
from nemo.collections import asr as nemo_asr
|
| 22 |
import gradio as gr
|
| 23 |
-
import noisereduce as nr
|
| 24 |
|
| 25 |
# ----------------------------
|
| 26 |
# CONFIG
|
|
@@ -53,49 +53,32 @@ def run_cmd(cmd):
|
|
| 53 |
return res.stdout
|
| 54 |
|
| 55 |
def ffprobe_duration(path):
|
| 56 |
-
"""
|
| 57 |
-
Tente d'obtenir la durée via ffprobe (robuste pour les conteneurs webcam).
|
| 58 |
-
"""
|
| 59 |
-
# FIXE: On lit la durée du format (conteneur)
|
| 60 |
cmd = f'ffprobe -v error -show_entries format=duration -of default=noprint_wrappers=1:nokey=1 {shlex.quote(path)}'
|
| 61 |
out = subprocess.run(cmd, shell=True, stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True)
|
| 62 |
-
|
| 63 |
-
if out.returncode != 0:
|
| 64 |
-
return None
|
| 65 |
try:
|
| 66 |
output = out.stdout.strip().split('\n')[0]
|
| 67 |
return float(output)
|
| 68 |
-
except:
|
| 69 |
-
return None
|
| 70 |
|
| 71 |
# ----------------------------
|
| 72 |
# LOAD MODEL (robust)
|
| 73 |
# ----------------------------
|
| 74 |
def load_model(name):
|
| 75 |
-
"""Charge le modèle NeMo correct
|
| 76 |
-
if name in _cache:
|
| 77 |
-
return _cache[name]
|
| 78 |
-
|
| 79 |
repo, mode = MODELS[name]
|
| 80 |
print(f"[LOAD] snapshot_download {repo} ...")
|
| 81 |
folder = snapshot_download(repo, local_dir_use_symlinks=False)
|
| 82 |
nemo_file = next((os.path.join(folder, f) for f in os.listdir(folder) if f.endswith(".nemo")), None)
|
| 83 |
-
if not nemo_file:
|
| 84 |
-
|
| 85 |
-
|
| 86 |
-
|
| 87 |
-
|
| 88 |
-
|
| 89 |
-
|
| 90 |
-
model = nemo_asr.models.EncDecHybridRNNTCTCBPEModel.restore_from(nemo_file)
|
| 91 |
-
elif mode == "ctc_char":
|
| 92 |
-
model = nemo_asr.models.EncDecCTCModel.restore_from(nemo_file)
|
| 93 |
-
else: # mode == "ctc" (BPE)
|
| 94 |
-
try:
|
| 95 |
-
model = nemo_asr.models.EncDecCTCModelBPE.restore_from(nemo_file)
|
| 96 |
-
except Exception as e:
|
| 97 |
-
print(f"[WARN] EncDecCTCModelBPE failed ({e}), fallback EncDecCTCModel")
|
| 98 |
-
model = nemo_asr.models.EncDecCTCModel.restore_from(nemo_file)
|
| 99 |
|
| 100 |
model.to(DEVICE).eval()
|
| 101 |
_cache[name] = model
|
|
@@ -111,83 +94,41 @@ def extract_audio(video_path, out_wav):
|
|
| 111 |
run_cmd(cmd)
|
| 112 |
|
| 113 |
def clean_audio(wav_path, target_sr=16000):
|
| 114 |
-
"""
|
| 115 |
-
Load audio, apply noise reduction, resample, normalize, write cleaned wav.
|
| 116 |
-
"""
|
| 117 |
audio, sr = sf.read(wav_path)
|
| 118 |
-
|
| 119 |
-
if audio.ndim == 2:
|
| 120 |
-
audio = audio.mean(axis=1)
|
| 121 |
-
|
| 122 |
if sr != target_sr:
|
| 123 |
audio = librosa.resample(audio.astype(float), orig_sr=sr, target_sr=target_sr)
|
| 124 |
sr = target_sr
|
| 125 |
-
|
| 126 |
-
# --- AMÉLIORATION : Réduction de bruit ---
|
| 127 |
try:
|
| 128 |
print("[INFO] Application de la réduction de bruit (noisereduce)...")
|
| 129 |
-
# Réduit le bruit stationnaire (ventilateur, souffle) de 75%
|
| 130 |
audio = nr.reduce_noise(y=audio, sr=sr, stationary=True, prop_decrease=0.75)
|
| 131 |
except Exception as e:
|
| 132 |
print(f"[WARN] Echec noisereduce: {e}")
|
| 133 |
-
|
| 134 |
-
|
| 135 |
max_val = np.max(np.abs(audio)) if audio.size > 0 else 0.0
|
| 136 |
-
if max_val > 1e-6:
|
| 137 |
-
# Normalisation à 0.95
|
| 138 |
-
audio = audio / max_val * 0.95
|
| 139 |
|
| 140 |
clean_path = str(Path(wav_path).with_name(Path(wav_path).stem + "_clean.wav"))
|
| 141 |
sf.write(clean_path, audio, sr)
|
| 142 |
-
|
| 143 |
return clean_path, audio, sr
|
| 144 |
|
| 145 |
# ----------------------------
|
| 146 |
-
# TRANSCRIPTION
|
| 147 |
# ----------------------------
|
| 148 |
def transcribe(model, wav_path):
|
| 149 |
"""Robuste: essaie model.transcribe et nettoie la sortie."""
|
| 150 |
-
if not hasattr(model, "transcribe"):
|
| 151 |
-
raise RuntimeError("Le modèle ne supporte pas model.transcribe()")
|
| 152 |
out = model.transcribe([wav_path])
|
| 153 |
-
if isinstance(out, list):
|
| 154 |
-
|
| 155 |
-
return ""
|
| 156 |
-
first = out[0]
|
| 157 |
-
if isinstance(first, str):
|
| 158 |
-
return first.strip()
|
| 159 |
-
if hasattr(first, "text"):
|
| 160 |
-
return first.text.strip()
|
| 161 |
-
return str(first).strip()
|
| 162 |
-
if hasattr(out, "text"):
|
| 163 |
-
return out.text.strip()
|
| 164 |
return str(out).strip()
|
| 165 |
|
| 166 |
-
# ----------------------------
|
| 167 |
-
# UTILITAIRES sous-titres / packing
|
| 168 |
-
# ----------------------------
|
| 169 |
-
def keep_bambara(words):
|
| 170 |
-
res = []
|
| 171 |
-
for w in words:
|
| 172 |
-
wl = w.lower()
|
| 173 |
-
if any(c in wl for c in ["ɛ","ɔ","ŋ"]) or sum(1 for c in wl if c in "aeiou") >= 2:
|
| 174 |
-
res.append(w)
|
| 175 |
-
return res
|
| 176 |
-
|
| 177 |
MAX_CHARS = 45; MIN_DUR = 0.3; MAX_DUR = 3.2; MAX_WORDS = 8
|
| 178 |
|
| 179 |
-
def wrap2(txt):
|
| 180 |
-
parts = textwrap.wrap(txt, MAX_CHARS)
|
| 181 |
-
if len(parts) <= 1:
|
| 182 |
-
return txt
|
| 183 |
-
mid = len(txt) // 2
|
| 184 |
-
left = txt.rfind(" ", 0, mid)
|
| 185 |
-
right = txt.find(" ", mid)
|
| 186 |
-
cut = left if (mid - left) <= ((right - mid) if right != -1 else 1e9) else right
|
| 187 |
-
l1 = txt[:cut].strip(); l2 = txt[cut:].strip()
|
| 188 |
-
return l1 + "\n" + l2 if l2 else l1
|
| 189 |
-
|
| 190 |
def pack(spans, total):
|
|
|
|
| 191 |
tmp = []
|
| 192 |
for s, e, t in spans:
|
| 193 |
s = max(0, min(s, total)); e = max(0, min(e, total))
|
|
@@ -211,34 +152,29 @@ def pack(spans, total):
|
|
| 211 |
for b in blocks:
|
| 212 |
st = base; en = min(base + step, e); base = en
|
| 213 |
if en <= st: en = min(st + 0.05, total)
|
| 214 |
-
txt =
|
|
|
|
| 215 |
if st < last_end:
|
| 216 |
st = last_end + 1e-3; en = max(en, st + 0.05)
|
| 217 |
out.append((st, en, txt)); last_end = en
|
| 218 |
return out
|
| 219 |
|
| 220 |
-
# ----------------------------
|
| 221 |
-
# VAD ALIGN (fallback alignment)
|
| 222 |
-
# ----------------------------
|
| 223 |
def align_vad(text, audio, sr, total_dur, top_db=28):
|
| 224 |
-
|
|
|
|
| 225 |
total = total_dur
|
| 226 |
if audio is None or len(audio) == 0 or not words:
|
| 227 |
return pack([(0, total, " ".join(words[:MAX_WORDS]))], total)
|
| 228 |
iv = librosa.effects.split(audio, top_db=top_db)
|
| 229 |
if len(iv) == 0:
|
| 230 |
return pack([(0, total, " ".join(words[:MAX_WORDS]))], total)
|
| 231 |
-
spans = []
|
| 232 |
-
L = sum(e - s for s, e in iv)
|
| 233 |
-
idx = 0
|
| 234 |
for s, e in iv:
|
| 235 |
seg = e - s; segt = seg / sr
|
| 236 |
-
k = max(1, int(round(len(words) * (seg / L))))
|
| 237 |
-
chunk = words[idx:idx+k]; idx += k
|
| 238 |
if not chunk: continue
|
| 239 |
lines = [chunk[i:i+MAX_WORDS] for i in range(0, len(chunk), MAX_WORDS)]
|
| 240 |
-
step = max(MIN_DUR, min(MAX_DUR, segt / max(1, len(lines))))
|
| 241 |
-
base = s / sr
|
| 242 |
for j, ln in enumerate(lines):
|
| 243 |
st = base + j * step; en = base + (j + 1) * step
|
| 244 |
spans.append((st, en, " ".join(ln)))
|
|
@@ -248,114 +184,70 @@ def align_vad(text, audio, sr, total_dur, top_db=28):
|
|
| 248 |
# Écriture SRT + Burn (réencode)
|
| 249 |
# ----------------------------
|
| 250 |
def burn(video_path, subs, output_path=None):
|
| 251 |
-
|
| 252 |
-
|
| 253 |
-
tmp_fd, tmp_srt = tempfile.mkstemp(suffix=".srt")
|
| 254 |
-
|
| 255 |
def sec_to_srt(t):
|
| 256 |
h = int(t // 3600); m = int((t % 3600) // 60); s = int(t % 60); ms = int((t - int(t)) * 1000)
|
| 257 |
return f"{h:02}:{m:02}:{s:02},{ms:03}"
|
|
|
|
| 258 |
with open(tmp_srt, "w", encoding="utf-8") as f:
|
| 259 |
for i, (start, end, text) in enumerate(subs, 1):
|
| 260 |
f.write(f"{i}\n{sec_to_srt(start)} --> {sec_to_srt(end)}\n{text}\n\n")
|
| 261 |
|
| 262 |
-
# On réencode (libx264) car on applique subtitles filter
|
| 263 |
vf = f"subtitles={shlex.quote(tmp_srt)}:force_style='Fontsize=22,PrimaryColour=&HFFFFFF&,OutlineColour=&H000000&'"
|
| 264 |
cmd = f'ffmpeg -hide_banner -loglevel error -y -i {shlex.quote(video_path)} -vf {shlex.quote(vf)} -c:v libx264 -preset fast -crf 23 -c:a aac -b:a 192k {shlex.quote(output_path)}'
|
| 265 |
-
try:
|
| 266 |
-
run_cmd(cmd)
|
| 267 |
finally:
|
| 268 |
-
if os.path.exists(tmp_srt):
|
| 269 |
-
os.remove(tmp_srt)
|
| 270 |
return output_path
|
| 271 |
|
| 272 |
# ----------------------------
|
| 273 |
# PIPELINE PRINCIPAL (Robuste)
|
| 274 |
# ----------------------------
|
| 275 |
def pipeline(video_input, model_name):
|
| 276 |
-
"""
|
| 277 |
-
video_input : chemin ou dict Gradio (tmp_path)
|
| 278 |
-
model_name : clé dans MODELS
|
| 279 |
-
"""
|
| 280 |
try:
|
| 281 |
-
if isinstance(video_input, dict) and "tmp_path" in video_input:
|
| 282 |
-
|
| 283 |
-
else:
|
| 284 |
-
video_path = video_input
|
| 285 |
|
| 286 |
-
# 1. Tentative d'obtention de durée via FFPROBE
|
| 287 |
duration = ffprobe_duration(video_path)
|
| 288 |
-
|
| 289 |
-
# 2. Extraction & Nettoyage Audio
|
| 290 |
-
tmp_fd, tmp_wav = tempfile.mkstemp(suffix=".wav")
|
| 291 |
-
os.close(tmp_fd)
|
| 292 |
-
|
| 293 |
extract_audio(video_path, tmp_wav)
|
| 294 |
clean_wav, audio, sr = clean_audio(tmp_wav)
|
| 295 |
|
| 296 |
-
# 3. FALLBACK: Si FFprobe a échoué (None), on calcule depuis l'audio
|
| 297 |
if duration is None:
|
| 298 |
print("[INFO] ffprobe duration failed, calculating from audio...")
|
| 299 |
-
if sr and sr > 0:
|
| 300 |
-
duration = len(audio) / sr
|
| 301 |
|
| 302 |
-
# Vérification finale
|
| 303 |
if not duration or duration <= 0:
|
| 304 |
raise RuntimeError("Impossible de déterminer la durée de la vidéo (fichier corrompu ?)")
|
| 305 |
|
| 306 |
-
print(f"[INFO] Durée détectée: {duration:.2f}s")
|
| 307 |
-
|
| 308 |
-
# 4. Chargement modèle + Transcription
|
| 309 |
model = load_model(model_name)
|
| 310 |
text = transcribe(model, clean_wav)
|
| 311 |
mode = MODELS[model_name][1]
|
| 312 |
|
| 313 |
-
#
|
| 314 |
-
subs = None
|
| 315 |
-
|
| 316 |
if mode == "rnnt":
|
| 317 |
-
# Logique d'alignement RNNT (CTC Segmentation)
|
| 318 |
try:
|
| 319 |
from ctc_segmentation import ctc_segmentation, CtcSegmentationParameters, prepare_text
|
| 320 |
-
words =
|
| 321 |
-
if not words:
|
| 322 |
-
|
| 323 |
-
|
| 324 |
-
ln = torch.tensor([x.shape[1]]).to(DEVICE)
|
| 325 |
-
with torch.no_grad():
|
| 326 |
-
logits = model(input_signal=x, input_signal_length=ln)[0]
|
| 327 |
time_per_frame = duration / max(1, logits.shape[1])
|
| 328 |
-
|
| 329 |
-
raw = model.tokenizer.vocab
|
| 330 |
-
vocab = list(raw.keys()) if isinstance(raw, dict) else list(raw)
|
| 331 |
-
except Exception:
|
| 332 |
-
vocab = None
|
| 333 |
-
cfg = CtcSegmentationParameters()
|
| 334 |
-
if vocab:
|
| 335 |
-
cfg.char_list = vocab
|
| 336 |
gt = prepare_text(cfg, words)[0]
|
| 337 |
-
|
| 338 |
-
|
| 339 |
-
|
| 340 |
-
|
| 341 |
-
except AssertionError:
|
| 342 |
-
print("[WARN] Audio shorter than text -> fallback to VAD alignment")
|
| 343 |
-
subs = align_vad(text, audio, sr, duration)
|
| 344 |
-
except Exception as e:
|
| 345 |
-
print(f"[WARN] ctc_segmentation not available or failed ({e}) -> fallback VAD")
|
| 346 |
-
subs = align_vad(text, audio, sr, duration)
|
| 347 |
-
|
| 348 |
-
elif mode == "ctc_char" or mode == "ctc":
|
| 349 |
-
# Logique d'alignement CTC / CTC-Char (VAD fallback)
|
| 350 |
-
try:
|
| 351 |
subs = align_vad(text, audio, sr, duration)
|
| 352 |
-
|
| 353 |
-
|
| 354 |
-
subs = align_vad(text, audio, sr, duration)
|
| 355 |
-
|
| 356 |
-
if not subs:
|
| 357 |
-
return ("⚠️ Aucun sous-titre utilisable (sub list vide)", None)
|
| 358 |
|
|
|
|
| 359 |
out_video = burn(video_path, subs)
|
| 360 |
return ("✅ Terminé avec succès", out_video)
|
| 361 |
|
|
@@ -364,13 +256,12 @@ def pipeline(video_input, model_name):
|
|
| 364 |
return (f"❌ Erreur — {str(e)}", None)
|
| 365 |
|
| 366 |
# ----------------------------
|
| 367 |
-
# INTERFACE GRADIO (
|
| 368 |
# ----------------------------
|
| 369 |
with gr.Blocks(title="RobotsMali - Sous-titrage") as demo:
|
| 370 |
gr.Markdown("## 🤖 RobotsMali — Sous-titrage Bambara (Amélioration Audio)")
|
| 371 |
|
| 372 |
# 1. Définir toutes les sorties AVANT leur utilisation.
|
| 373 |
-
# Elles sont rendues ici implicitement et sont disponibles pour gr.Examples.
|
| 374 |
s = gr.Markdown(label="Statut de la tâche")
|
| 375 |
o = gr.Video(label="Vidéo sous-titrée")
|
| 376 |
|
|
@@ -380,31 +271,30 @@ with gr.Blocks(title="RobotsMali - Sous-titrage") as demo:
|
|
| 380 |
v = gr.Video(label="Vidéo à sous-titrer", sources=["upload", "webcam"])
|
| 381 |
m = gr.Dropdown(list(MODELS.keys()), value="Soloba V1 (CTC)", label="Modèle ASR")
|
| 382 |
|
| 383 |
-
# 3. gr.Examples
|
| 384 |
gr.Examples(
|
| 385 |
examples=[
|
| 386 |
-
|
|
|
|
| 387 |
],
|
| 388 |
inputs=[v, m],
|
| 389 |
fn=pipeline,
|
| 390 |
outputs=[s, o],
|
| 391 |
label="▶️ Utiliser un exemple (Vidéo stockée dans le Space)",
|
| 392 |
-
run_on_click=True
|
|
|
|
| 393 |
)
|
| 394 |
|
| 395 |
b = gr.Button("▶️ Générer les sous-titres", variant="primary")
|
| 396 |
|
| 397 |
with gr.Column():
|
| 398 |
-
# 4.
|
| 399 |
-
# et on fait confiance à Gradio pour afficher S et O dans l'ordre de leur définition.
|
| 400 |
gr.Markdown("### Résultats:")
|
| 401 |
-
|
| 402 |
-
|
| 403 |
-
# Il n'y a rien à faire ici, à part s'assurer qu'ils sont bien affichés
|
| 404 |
-
# (ce qui est le cas par leur définition initiale dans le bloc).
|
| 405 |
|
| 406 |
-
# 5.
|
| 407 |
b.click(pipeline, [v, m], [s, o])
|
| 408 |
|
| 409 |
if __name__ == "__main__":
|
| 410 |
-
demo.launch(share=True
|
|
|
|
| 20 |
from huggingface_hub import snapshot_download
|
| 21 |
from nemo.collections import asr as nemo_asr
|
| 22 |
import gradio as gr
|
| 23 |
+
import noisereduce as nr
|
| 24 |
|
| 25 |
# ----------------------------
|
| 26 |
# CONFIG
|
|
|
|
| 53 |
return res.stdout
|
| 54 |
|
| 55 |
def ffprobe_duration(path):
|
| 56 |
+
"""Tente d'obtenir la durée via ffprobe."""
|
|
|
|
|
|
|
|
|
|
| 57 |
cmd = f'ffprobe -v error -show_entries format=duration -of default=noprint_wrappers=1:nokey=1 {shlex.quote(path)}'
|
| 58 |
out = subprocess.run(cmd, shell=True, stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True)
|
| 59 |
+
if out.returncode != 0: return None
|
|
|
|
|
|
|
| 60 |
try:
|
| 61 |
output = out.stdout.strip().split('\n')[0]
|
| 62 |
return float(output)
|
| 63 |
+
except: return None
|
|
|
|
| 64 |
|
| 65 |
# ----------------------------
|
| 66 |
# LOAD MODEL (robust)
|
| 67 |
# ----------------------------
|
| 68 |
def load_model(name):
|
| 69 |
+
"""Charge le modèle NeMo correct."""
|
| 70 |
+
if name in _cache: return _cache[name]
|
|
|
|
|
|
|
| 71 |
repo, mode = MODELS[name]
|
| 72 |
print(f"[LOAD] snapshot_download {repo} ...")
|
| 73 |
folder = snapshot_download(repo, local_dir_use_symlinks=False)
|
| 74 |
nemo_file = next((os.path.join(folder, f) for f in os.listdir(folder) if f.endswith(".nemo")), None)
|
| 75 |
+
if not nemo_file: raise FileNotFoundError(f"Aucun .nemo trouvé pour {name} dans {folder}")
|
| 76 |
+
|
| 77 |
+
if mode == "rnnt": model = nemo_asr.models.EncDecHybridRNNTCTCBPEModel.restore_from(nemo_file)
|
| 78 |
+
elif mode == "ctc_char": model = nemo_asr.models.EncDecCTCModel.restore_from(nemo_file)
|
| 79 |
+
else:
|
| 80 |
+
try: model = nemo_asr.models.EncDecCTCModelBPE.restore_from(nemo_file)
|
| 81 |
+
except Exception: model = nemo_asr.models.EncDecCTCModel.restore_from(nemo_file)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 82 |
|
| 83 |
model.to(DEVICE).eval()
|
| 84 |
_cache[name] = model
|
|
|
|
| 94 |
run_cmd(cmd)
|
| 95 |
|
| 96 |
def clean_audio(wav_path, target_sr=16000):
|
| 97 |
+
"""Load audio, apply noise reduction, resample, normalize, write cleaned wav."""
|
|
|
|
|
|
|
| 98 |
audio, sr = sf.read(wav_path)
|
| 99 |
+
if audio.ndim == 2: audio = audio.mean(axis=1)
|
|
|
|
|
|
|
|
|
|
| 100 |
if sr != target_sr:
|
| 101 |
audio = librosa.resample(audio.astype(float), orig_sr=sr, target_sr=target_sr)
|
| 102 |
sr = target_sr
|
| 103 |
+
|
|
|
|
| 104 |
try:
|
| 105 |
print("[INFO] Application de la réduction de bruit (noisereduce)...")
|
|
|
|
| 106 |
audio = nr.reduce_noise(y=audio, sr=sr, stationary=True, prop_decrease=0.75)
|
| 107 |
except Exception as e:
|
| 108 |
print(f"[WARN] Echec noisereduce: {e}")
|
| 109 |
+
|
|
|
|
| 110 |
max_val = np.max(np.abs(audio)) if audio.size > 0 else 0.0
|
| 111 |
+
if max_val > 1e-6: audio = audio / max_val * 0.95
|
|
|
|
|
|
|
| 112 |
|
| 113 |
clean_path = str(Path(wav_path).with_name(Path(wav_path).stem + "_clean.wav"))
|
| 114 |
sf.write(clean_path, audio, sr)
|
|
|
|
| 115 |
return clean_path, audio, sr
|
| 116 |
|
| 117 |
# ----------------------------
|
| 118 |
+
# TRANSCRIPTION & ALIGNMENT UTILS
|
| 119 |
# ----------------------------
|
| 120 |
def transcribe(model, wav_path):
|
| 121 |
"""Robuste: essaie model.transcribe et nettoie la sortie."""
|
| 122 |
+
if not hasattr(model, "transcribe"): raise RuntimeError("Le modèle ne supporte pas model.transcribe()")
|
|
|
|
| 123 |
out = model.transcribe([wav_path])
|
| 124 |
+
if isinstance(out, list) and len(out) > 0: out = out[0]
|
| 125 |
+
if hasattr(out, "text"): return out.text.strip()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 126 |
return str(out).strip()
|
| 127 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 128 |
MAX_CHARS = 45; MIN_DUR = 0.3; MAX_DUR = 3.2; MAX_WORDS = 8
|
| 129 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 130 |
def pack(spans, total):
|
| 131 |
+
# Logique complexe de regroupement et de réemballage (non modifiée)
|
| 132 |
tmp = []
|
| 133 |
for s, e, t in spans:
|
| 134 |
s = max(0, min(s, total)); e = max(0, min(e, total))
|
|
|
|
| 152 |
for b in blocks:
|
| 153 |
st = base; en = min(base + step, e); base = en
|
| 154 |
if en <= st: en = min(st + 0.05, total)
|
| 155 |
+
txt = textwrap.wrap(b, MAX_CHARS)
|
| 156 |
+
txt = txt[0] + "\n" + txt[1] if len(txt) > 1 else txt[0]
|
| 157 |
if st < last_end:
|
| 158 |
st = last_end + 1e-3; en = max(en, st + 0.05)
|
| 159 |
out.append((st, en, txt)); last_end = en
|
| 160 |
return out
|
| 161 |
|
|
|
|
|
|
|
|
|
|
| 162 |
def align_vad(text, audio, sr, total_dur, top_db=28):
|
| 163 |
+
# Logique VAD (non modifiée)
|
| 164 |
+
words = [w for w in text.split() if any(c in w.lower() for c in ["ɛ","ɔ","ŋ"]) or sum(1 for c in w.lower() if c in "aeiou") >= 2]
|
| 165 |
total = total_dur
|
| 166 |
if audio is None or len(audio) == 0 or not words:
|
| 167 |
return pack([(0, total, " ".join(words[:MAX_WORDS]))], total)
|
| 168 |
iv = librosa.effects.split(audio, top_db=top_db)
|
| 169 |
if len(iv) == 0:
|
| 170 |
return pack([(0, total, " ".join(words[:MAX_WORDS]))], total)
|
| 171 |
+
spans = []; L = sum(e - s for s, e in iv); idx = 0
|
|
|
|
|
|
|
| 172 |
for s, e in iv:
|
| 173 |
seg = e - s; segt = seg / sr
|
| 174 |
+
k = max(1, int(round(len(words) * (seg / L)))); chunk = words[idx:idx+k]; idx += k
|
|
|
|
| 175 |
if not chunk: continue
|
| 176 |
lines = [chunk[i:i+MAX_WORDS] for i in range(0, len(chunk), MAX_WORDS)]
|
| 177 |
+
step = max(MIN_DUR, min(MAX_DUR, segt / max(1, len(lines)))); base = s / sr
|
|
|
|
| 178 |
for j, ln in enumerate(lines):
|
| 179 |
st = base + j * step; en = base + (j + 1) * step
|
| 180 |
spans.append((st, en, " ".join(ln)))
|
|
|
|
| 184 |
# Écriture SRT + Burn (réencode)
|
| 185 |
# ----------------------------
|
| 186 |
def burn(video_path, subs, output_path=None):
|
| 187 |
+
"""Crée le SRT temporaire et brûle les sous-titres dans la vidéo."""
|
| 188 |
+
if output_path is None: output_path = "RobotsMali_Subtitled.mp4"
|
| 189 |
+
tmp_fd, tmp_srt = tempfile.mkstemp(suffix=".srt"); os.close(tmp_fd)
|
| 190 |
+
|
| 191 |
def sec_to_srt(t):
|
| 192 |
h = int(t // 3600); m = int((t % 3600) // 60); s = int(t % 60); ms = int((t - int(t)) * 1000)
|
| 193 |
return f"{h:02}:{m:02}:{s:02},{ms:03}"
|
| 194 |
+
|
| 195 |
with open(tmp_srt, "w", encoding="utf-8") as f:
|
| 196 |
for i, (start, end, text) in enumerate(subs, 1):
|
| 197 |
f.write(f"{i}\n{sec_to_srt(start)} --> {sec_to_srt(end)}\n{text}\n\n")
|
| 198 |
|
|
|
|
| 199 |
vf = f"subtitles={shlex.quote(tmp_srt)}:force_style='Fontsize=22,PrimaryColour=&HFFFFFF&,OutlineColour=&H000000&'"
|
| 200 |
cmd = f'ffmpeg -hide_banner -loglevel error -y -i {shlex.quote(video_path)} -vf {shlex.quote(vf)} -c:v libx264 -preset fast -crf 23 -c:a aac -b:a 192k {shlex.quote(output_path)}'
|
| 201 |
+
try: run_cmd(cmd)
|
|
|
|
| 202 |
finally:
|
| 203 |
+
if os.path.exists(tmp_srt): os.remove(tmp_srt)
|
|
|
|
| 204 |
return output_path
|
| 205 |
|
| 206 |
# ----------------------------
|
| 207 |
# PIPELINE PRINCIPAL (Robuste)
|
| 208 |
# ----------------------------
|
| 209 |
def pipeline(video_input, model_name):
|
| 210 |
+
"""Gère le flux de sous-titrage complet."""
|
|
|
|
|
|
|
|
|
|
| 211 |
try:
|
| 212 |
+
if isinstance(video_input, dict) and "tmp_path" in video_input: video_path = video_input["tmp_path"]
|
| 213 |
+
else: video_path = video_input
|
|
|
|
|
|
|
| 214 |
|
|
|
|
| 215 |
duration = ffprobe_duration(video_path)
|
| 216 |
+
tmp_fd, tmp_wav = tempfile.mkstemp(suffix=".wav"); os.close(tmp_fd)
|
|
|
|
|
|
|
|
|
|
|
|
|
| 217 |
extract_audio(video_path, tmp_wav)
|
| 218 |
clean_wav, audio, sr = clean_audio(tmp_wav)
|
| 219 |
|
|
|
|
| 220 |
if duration is None:
|
| 221 |
print("[INFO] ffprobe duration failed, calculating from audio...")
|
| 222 |
+
if sr and sr > 0: duration = len(audio) / sr
|
|
|
|
| 223 |
|
|
|
|
| 224 |
if not duration or duration <= 0:
|
| 225 |
raise RuntimeError("Impossible de déterminer la durée de la vidéo (fichier corrompu ?)")
|
| 226 |
|
|
|
|
|
|
|
|
|
|
| 227 |
model = load_model(model_name)
|
| 228 |
text = transcribe(model, clean_wav)
|
| 229 |
mode = MODELS[model_name][1]
|
| 230 |
|
| 231 |
+
# Logique d'alignement (CTC Segmentation ou VAD Fallback)
|
|
|
|
|
|
|
| 232 |
if mode == "rnnt":
|
|
|
|
| 233 |
try:
|
| 234 |
from ctc_segmentation import ctc_segmentation, CtcSegmentationParameters, prepare_text
|
| 235 |
+
words = [w for w in text.split() if any(c in w.lower() for c in ["ɛ","ɔ","ŋ"]) or sum(1 for c in w.lower() if c in "aeiou") >= 2]
|
| 236 |
+
if not words: return ("⚠️ Aucun sous-titre utilisable (texte vide après filtrage)", None)
|
| 237 |
+
x = torch.tensor(audio).float().unsqueeze(0).to(DEVICE); ln = torch.tensor([x.shape[1]]).to(DEVICE)
|
| 238 |
+
with torch.no_grad(): logits = model(input_signal=x, input_signal_length=ln)[0]
|
|
|
|
|
|
|
|
|
|
| 239 |
time_per_frame = duration / max(1, logits.shape[1])
|
| 240 |
+
cfg = CtcSegmentationParameters(); cfg.char_list = list(model.tokenizer.vocab.keys())
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 241 |
gt = prepare_text(cfg, words)[0]
|
| 242 |
+
timing, _, _ = ctc_segmentation(cfg, logits.detach().cpu().numpy()[0], gt)
|
| 243 |
+
spans = [(timing[i] * time_per_frame, timing[i+1] * time_per_frame, words[i]) for i in range(len(words) - 1)]
|
| 244 |
+
subs = pack(spans, duration)
|
| 245 |
+
except Exception:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 246 |
subs = align_vad(text, audio, sr, duration)
|
| 247 |
+
else:
|
| 248 |
+
subs = align_vad(text, audio, sr, duration)
|
|
|
|
|
|
|
|
|
|
|
|
|
| 249 |
|
| 250 |
+
if not subs: return ("⚠️ Aucun sous-titre utilisable (sub list vide)", None)
|
| 251 |
out_video = burn(video_path, subs)
|
| 252 |
return ("✅ Terminé avec succès", out_video)
|
| 253 |
|
|
|
|
| 256 |
return (f"❌ Erreur — {str(e)}", None)
|
| 257 |
|
| 258 |
# ----------------------------
|
| 259 |
+
# INTERFACE GRADIO (Version Finale Stabilité)
|
| 260 |
# ----------------------------
|
| 261 |
with gr.Blocks(title="RobotsMali - Sous-titrage") as demo:
|
| 262 |
gr.Markdown("## 🤖 RobotsMali — Sous-titrage Bambara (Amélioration Audio)")
|
| 263 |
|
| 264 |
# 1. Définir toutes les sorties AVANT leur utilisation.
|
|
|
|
| 265 |
s = gr.Markdown(label="Statut de la tâche")
|
| 266 |
o = gr.Video(label="Vidéo sous-titrée")
|
| 267 |
|
|
|
|
| 271 |
v = gr.Video(label="Vidéo à sous-titrer", sources=["upload", "webcam"])
|
| 272 |
m = gr.Dropdown(list(MODELS.keys()), value="Soloba V1 (CTC)", label="Modèle ASR")
|
| 273 |
|
| 274 |
+
# 3. gr.Examples (avec cache_examples=False et nom de fichier corrigé)
|
| 275 |
gr.Examples(
|
| 276 |
examples=[
|
| 277 |
+
# Utiliser le nom de fichier exact du dépôt
|
| 278 |
+
["examples/Upload MARALINKE-WILI (Lève-toi) Black lives matter (Clip officiel) - MARALINKE (360p, h264).mp4", "Soloba V1 (CTC)"]
|
| 279 |
],
|
| 280 |
inputs=[v, m],
|
| 281 |
fn=pipeline,
|
| 282 |
outputs=[s, o],
|
| 283 |
label="▶️ Utiliser un exemple (Vidéo stockée dans le Space)",
|
| 284 |
+
run_on_click=True,
|
| 285 |
+
cache_examples=False
|
| 286 |
)
|
| 287 |
|
| 288 |
b = gr.Button("▶️ Générer les sous-titres", variant="primary")
|
| 289 |
|
| 290 |
with gr.Column():
|
| 291 |
+
# 4. Affichage des sorties
|
|
|
|
| 292 |
gr.Markdown("### Résultats:")
|
| 293 |
+
s
|
| 294 |
+
o
|
|
|
|
|
|
|
| 295 |
|
| 296 |
+
# 5. L'action du bouton
|
| 297 |
b.click(pipeline, [v, m], [s, o])
|
| 298 |
|
| 299 |
if __name__ == "__main__":
|
| 300 |
+
demo.launch(share=True)
|