Update app.py
Browse files
app.py
CHANGED
|
@@ -1,8 +1,21 @@
|
|
| 1 |
# -*- coding: utf-8 -*-
|
| 2 |
-
"""
|
| 3 |
-
|
| 4 |
-
|
| 5 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 6 |
|
| 7 |
import os
|
| 8 |
import shlex
|
|
@@ -12,7 +25,7 @@ import traceback
|
|
| 12 |
import random
|
| 13 |
import textwrap
|
| 14 |
from pathlib import Path
|
| 15 |
-
|
| 16 |
import numpy as np
|
| 17 |
import torch
|
| 18 |
import soundfile as sf
|
|
@@ -20,55 +33,82 @@ 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 |
-
|
| 27 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 28 |
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
|
| 29 |
random.seed(1234)
|
| 30 |
np.random.seed(1234)
|
| 31 |
torch.manual_seed(1234)
|
| 32 |
|
|
|
|
|
|
|
|
|
|
| 33 |
MODELS = {
|
| 34 |
"Soloni V1 (RNNT)": ("RobotsMali/soloni-114m-tdt-ctc-v1", "rnnt"),
|
| 35 |
"Soloni V0 (RNNT)": ("RobotsMali/soloni-114m-tdt-ctc-v0", "rnnt"),
|
| 36 |
-
"Soloba V1 (CTC)": ("RobotsMali/soloba-ctc-0.6b-v1", "ctc"),
|
| 37 |
-
"Soloba V0 (CTC)": ("RobotsMali/soloba-ctc-0.6b-v0", "ctc"),
|
| 38 |
-
"QuartzNet V1 (CTC-char)": ("RobotsMali/stt-bm-quartznet15x5-v1", "ctc_char"),
|
| 39 |
"QuartzNet V0 (CTC-char)": ("RobotsMali/stt-bm-quartznet15x5-v0", "ctc_char"),
|
| 40 |
}
|
| 41 |
-
|
| 42 |
_cache = {}
|
| 43 |
|
| 44 |
-
#
|
| 45 |
-
|
| 46 |
-
|
|
|
|
|
|
|
|
|
|
| 47 |
def run_cmd(cmd):
|
|
|
|
|
|
|
| 48 |
res = subprocess.run(cmd, shell=True, stdout=subprocess.PIPE, stderr=subprocess.STDOUT, text=True)
|
| 49 |
if res.returncode != 0:
|
| 50 |
raise RuntimeError(f"Commande échouée [{cmd}]\nOutput:\n{res.stdout}")
|
| 51 |
return res.stdout
|
| 52 |
-
|
| 53 |
def ffprobe_duration(path):
|
|
|
|
| 54 |
cmd = f'ffprobe -v error -show_entries format=duration -of default=noprint_wrappers=1:nokey=1 {shlex.quote(path)}'
|
| 55 |
out = subprocess.run(cmd, shell=True, stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True)
|
| 56 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 57 |
try:
|
| 58 |
-
|
| 59 |
-
|
| 60 |
-
|
| 61 |
-
|
| 62 |
-
|
| 63 |
-
|
| 64 |
-
# ----------------------------
|
| 65 |
def load_model(name):
|
| 66 |
-
|
|
|
|
|
|
|
|
|
|
| 67 |
repo, mode = MODELS[name]
|
|
|
|
| 68 |
folder = snapshot_download(repo, local_dir_use_symlinks=False)
|
| 69 |
nemo_file = next((os.path.join(folder, f) for f in os.listdir(folder) if f.endswith(".nemo")), None)
|
| 70 |
if not nemo_file:
|
| 71 |
raise FileNotFoundError(f"Aucun .nemo trouvé pour {name} dans {folder}")
|
|
|
|
|
|
|
|
|
|
| 72 |
if mode == "rnnt":
|
| 73 |
model = nemo_asr.models.EncDecHybridRNNTCTCBPEModel.restore_from(nemo_file)
|
| 74 |
elif mode == "ctc_char":
|
|
@@ -76,97 +116,318 @@ def load_model(name):
|
|
| 76 |
else:
|
| 77 |
try:
|
| 78 |
model = nemo_asr.models.EncDecCTCModelBPE.restore_from(nemo_file)
|
| 79 |
-
except:
|
|
|
|
| 80 |
model = nemo_asr.models.EncDecCTCModel.restore_from(nemo_file)
|
|
|
|
| 81 |
model.to(DEVICE).eval()
|
| 82 |
_cache[name] = model
|
|
|
|
| 83 |
return model
|
| 84 |
-
|
| 85 |
-
# ----------------------------
|
| 86 |
-
# AUDIO EXTRACTION & CLEAN
|
| 87 |
-
# ----------------------------
|
| 88 |
def extract_audio(video_path, out_wav):
|
| 89 |
-
|
| 90 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 91 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 92 |
def clean_audio(wav_path, target_sr=16000):
|
|
|
|
| 93 |
audio, sr = sf.read(wav_path)
|
| 94 |
-
if audio.ndim == 2:
|
|
|
|
| 95 |
if sr != target_sr:
|
| 96 |
audio = librosa.resample(audio.astype(float), orig_sr=sr, target_sr=target_sr)
|
| 97 |
sr = target_sr
|
| 98 |
-
|
| 99 |
-
audio = nr.reduce_noise(y=audio, sr=sr, stationary=True, prop_decrease=0.75)
|
| 100 |
-
except: pass
|
| 101 |
-
max_val = np.max(np.abs(audio)) if audio.size > 0 else 0
|
| 102 |
if max_val > 1e-6:
|
| 103 |
-
audio = audio / max_val * 0.
|
| 104 |
clean_path = str(Path(wav_path).with_name(Path(wav_path).stem + "_clean.wav"))
|
| 105 |
sf.write(clean_path, audio, sr)
|
| 106 |
return clean_path, audio, sr
|
|
|
|
|
|
|
|
|
|
|
|
|
| 107 |
|
| 108 |
-
# ----------------------------
|
| 109 |
-
# TRANSCRIPTION
|
| 110 |
-
# ----------------------------
|
| 111 |
def transcribe(model, wav_path):
|
|
|
|
|
|
|
| 112 |
out = model.transcribe([wav_path])
|
| 113 |
-
if isinstance(out, list)
|
| 114 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 115 |
return str(out).strip()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 116 |
|
| 117 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 118 |
|
| 119 |
-
|
| 120 |
-
|
| 121 |
-
|
| 122 |
-
|
| 123 |
-
|
| 124 |
-
|
| 125 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 126 |
|
| 127 |
-
|
| 128 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 129 |
|
| 130 |
-
|
| 131 |
-
|
| 132 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 133 |
|
| 134 |
-
return
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 135 |
|
| 136 |
-
|
| 137 |
-
# INTERFACE GRADIO
|
| 138 |
-
# ----------------------------
|
| 139 |
with gr.Blocks(title="RobotsMali - Sous-titrage") as demo:
|
| 140 |
-
gr.Markdown("## 🤖 RobotsMali — Sous-titrage Bambara")
|
| 141 |
-
|
| 142 |
-
|
| 143 |
-
|
| 144 |
-
|
| 145 |
-
with gr.Row():
|
| 146 |
-
with gr.Column():
|
| 147 |
-
v = gr.Video(label="Vidéo à sous-titrer", sources=["upload", "webcam"])
|
| 148 |
-
m = gr.Dropdown(list(MODELS.keys()), value="Soloba V1 (CTC)", label="Modèle ASR")
|
| 149 |
-
|
| 150 |
-
gr.Examples(
|
| 151 |
-
examples=[
|
| 152 |
-
[get_example_video(), "Soloba V1 (CTC)"]
|
| 153 |
-
],
|
| 154 |
-
inputs=[v, m],
|
| 155 |
-
fn=pipeline,
|
| 156 |
-
outputs=[s, o],
|
| 157 |
-
label="▶️ Vidéo d’exemple du Space",
|
| 158 |
-
run_on_click=True,
|
| 159 |
-
cache_examples=False
|
| 160 |
-
)
|
| 161 |
-
|
| 162 |
-
b = gr.Button("▶️ Générer les sous-titres")
|
| 163 |
-
|
| 164 |
-
with gr.Column():
|
| 165 |
-
gr.Markdown("### Résultats :")
|
| 166 |
-
s
|
| 167 |
-
o
|
| 168 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 169 |
b.click(pipeline, [v, m], [s, o])
|
| 170 |
-
|
| 171 |
-
|
| 172 |
-
demo.launch(share=True, debug=True)
|
|
|
|
| 1 |
# -*- coding: utf-8 -*-
|
| 2 |
+
""" ROBOTSMALI — Sous-titrage Bambara (V4.8 Colab Ready - Remuxage Vidéo) """
|
| 3 |
+
|
| 4 |
+
# =========================================================================
|
| 5 |
+
# 1. INSTALLATION ET MISE À JOUR DES DÉPENDANCES
|
| 6 |
+
# =========================================================================
|
| 7 |
+
print("Démarrage de l'installation des dépendances...")
|
| 8 |
+
|
| 9 |
+
# Installation de FFmpeg (nécessaire pour ffprobe et extraction/burn)
|
| 10 |
+
!apt-get update && apt-get install -y ffmpeg
|
| 11 |
+
|
| 12 |
+
# Installation des librairies Python essentielles et des outils NeMo
|
| 13 |
+
!pip install torch numpy soundfile librosa huggingface_hub gradio
|
| 14 |
+
!pip install nemo_toolkit[asr]
|
| 15 |
+
!pip install ctc-segmentation
|
| 16 |
+
!pip install --upgrade gradio # Mise à jour de Gradio pour la compatibilité
|
| 17 |
+
print("Installation des dépendances terminée.")
|
| 18 |
+
# =========================================================================
|
| 19 |
|
| 20 |
import os
|
| 21 |
import shlex
|
|
|
|
| 25 |
import random
|
| 26 |
import textwrap
|
| 27 |
from pathlib import Path
|
| 28 |
+
|
| 29 |
import numpy as np
|
| 30 |
import torch
|
| 31 |
import soundfile as sf
|
|
|
|
| 33 |
from huggingface_hub import snapshot_download
|
| 34 |
from nemo.collections import asr as nemo_asr
|
| 35 |
import gradio as gr
|
|
|
|
| 36 |
|
| 37 |
+
# Tente l'importation de la librairie d'alignement nécessaire
|
| 38 |
+
try:
|
| 39 |
+
from ctc_segmentation import ctc_segmentation, CtcSegmentationParameters, prepare_text
|
| 40 |
+
HAS_CTC_SEGMENTATION = True
|
| 41 |
+
except ImportError:
|
| 42 |
+
HAS_CTC_SEGMENTATION = False
|
| 43 |
+
print("ATTENTION: ctc_segmentation non installé. L'alignement sera basé sur une simple répartition égale du temps.")
|
| 44 |
+
|
| 45 |
+
# ---------------------------- # CONFIG # ----------------------------
|
| 46 |
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
|
| 47 |
random.seed(1234)
|
| 48 |
np.random.seed(1234)
|
| 49 |
torch.manual_seed(1234)
|
| 50 |
|
| 51 |
+
# Taille du segment pour la transcription par blocs
|
| 52 |
+
SEGMENT_DURATION = 10.0
|
| 53 |
+
|
| 54 |
MODELS = {
|
| 55 |
"Soloni V1 (RNNT)": ("RobotsMali/soloni-114m-tdt-ctc-v1", "rnnt"),
|
| 56 |
"Soloni V0 (RNNT)": ("RobotsMali/soloni-114m-tdt-ctc-v0", "rnnt"),
|
| 57 |
+
"Soloba V1 (CTC)": ("RobotsMali/soloba-ctc-0.6b-v1", "ctc"),
|
| 58 |
+
"Soloba V0 (CTC)": ("RobotsMali/soloba-ctc-0.6b-v0", "ctc"),
|
| 59 |
+
"QuartzNet V1 (CTC-char)": ("RobotsMali/stt-bm-quartznet15x5-v1", "ctc_char"),
|
| 60 |
"QuartzNet V0 (CTC-char)": ("RobotsMali/stt-bm-quartznet15x5-v0", "ctc_char"),
|
| 61 |
}
|
| 62 |
+
|
| 63 |
_cache = {}
|
| 64 |
|
| 65 |
+
# Chemin vers la vidéo d'exemple.
|
| 66 |
+
VIDEO_EXAMPLES = [
|
| 67 |
+
"examples/MARALINKE-Wii (Lève-toi) Black lives matter (Clip officiel) - MARALINKE (360p, H264).mp4"
|
| 68 |
+
]
|
| 69 |
+
|
| 70 |
+
# ---------------------------- # UTIL: run_cmd, ffprobe_duration # ----------------------------
|
| 71 |
def run_cmd(cmd):
|
| 72 |
+
"""Execute a shell command and raise on non-zero exit."""
|
| 73 |
+
print("RUN:", cmd)
|
| 74 |
res = subprocess.run(cmd, shell=True, stdout=subprocess.PIPE, stderr=subprocess.STDOUT, text=True)
|
| 75 |
if res.returncode != 0:
|
| 76 |
raise RuntimeError(f"Commande échouée [{cmd}]\nOutput:\n{res.stdout}")
|
| 77 |
return res.stdout
|
| 78 |
+
|
| 79 |
def ffprobe_duration(path):
|
| 80 |
+
"""Détermine la durée de la vidéo via ffprobe (pour vérification/débogage)."""
|
| 81 |
cmd = f'ffprobe -v error -show_entries format=duration -of default=noprint_wrappers=1:nokey=1 {shlex.quote(path)}'
|
| 82 |
out = subprocess.run(cmd, shell=True, stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True)
|
| 83 |
+
|
| 84 |
+
if out.returncode != 0:
|
| 85 |
+
# Affiche l'erreur FFPROBE brute si l'extraction échoue
|
| 86 |
+
print(f"--- ERREUR FFPROBE BRUTE --- (Code: {out.returncode})")
|
| 87 |
+
print(out.stderr)
|
| 88 |
+
print("----------------------------")
|
| 89 |
+
return None
|
| 90 |
try:
|
| 91 |
+
return float(out.stdout.strip())
|
| 92 |
+
except Exception as e:
|
| 93 |
+
print(f"--- ERREUR CONVERSION DURÉE --- (Output: {out.stdout.strip()})")
|
| 94 |
+
print(e)
|
| 95 |
+
return None
|
| 96 |
+
|
| 97 |
+
# ---------------------------- # LOAD MODEL (robust) # ----------------------------
|
| 98 |
def load_model(name):
|
| 99 |
+
"""Charge le modèle NeMo correct selon type (rnnt / ctc / ctc_char)."""
|
| 100 |
+
if name in _cache:
|
| 101 |
+
return _cache[name]
|
| 102 |
+
|
| 103 |
repo, mode = MODELS[name]
|
| 104 |
+
print(f"[LOAD] snapshot_download {repo} ...")
|
| 105 |
folder = snapshot_download(repo, local_dir_use_symlinks=False)
|
| 106 |
nemo_file = next((os.path.join(folder, f) for f in os.listdir(folder) if f.endswith(".nemo")), None)
|
| 107 |
if not nemo_file:
|
| 108 |
raise FileNotFoundError(f"Aucun .nemo trouvé pour {name} dans {folder}")
|
| 109 |
+
|
| 110 |
+
print(f"[LOAD] .nemo trouvé: {nemo_file}; mode={mode}")
|
| 111 |
+
|
| 112 |
if mode == "rnnt":
|
| 113 |
model = nemo_asr.models.EncDecHybridRNNTCTCBPEModel.restore_from(nemo_file)
|
| 114 |
elif mode == "ctc_char":
|
|
|
|
| 116 |
else:
|
| 117 |
try:
|
| 118 |
model = nemo_asr.models.EncDecCTCModelBPE.restore_from(nemo_file)
|
| 119 |
+
except Exception as e:
|
| 120 |
+
print(f"[WARN] EncDecCTCModelBPE failed ({e}), fallback EncDecCTCModel")
|
| 121 |
model = nemo_asr.models.EncDecCTCModel.restore_from(nemo_file)
|
| 122 |
+
|
| 123 |
model.to(DEVICE).eval()
|
| 124 |
_cache[name] = model
|
| 125 |
+
print(f"[OK] Modèle {name} chargé sur {DEVICE}")
|
| 126 |
return model
|
| 127 |
+
|
| 128 |
+
# ---------------------------- # AUDIO EXTRACTION & CLEANING (ROBUSTE) # ----------------------------
|
|
|
|
|
|
|
| 129 |
def extract_audio(video_path, out_wav):
|
| 130 |
+
"""
|
| 131 |
+
Extrait l'audio en deux étapes pour stabiliser le fichier webcam/corrompu (Remuxage).
|
| 132 |
+
"""
|
| 133 |
+
# Chemin du fichier intermédiaire stabilisé
|
| 134 |
+
tmp_fd, stabilized_mp4 = tempfile.mkstemp(suffix="_stabilized.mp4")
|
| 135 |
+
os.close(tmp_fd)
|
| 136 |
|
| 137 |
+
# ÉTAPE 1: Stabilisation par Copie de Flux (Remuxage) du fichier d'entrée vers un conteneur MP4 stable.
|
| 138 |
+
# Ceci réécrit les métadonnées de l'en-tête (duration) sans réencoder.
|
| 139 |
+
remux_cmd = (
|
| 140 |
+
f'ffmpeg -hide_banner -loglevel error -y '
|
| 141 |
+
f'-analyzeduration 2147483647 -probesize 2147483647 -ignore_unknown '
|
| 142 |
+
f'-i {shlex.quote(video_path)} -c copy {shlex.quote(stabilized_mp4)}'
|
| 143 |
+
)
|
| 144 |
+
print("RUN: Remuxage du fichier pour stabilisation...")
|
| 145 |
+
run_cmd(remux_cmd)
|
| 146 |
+
|
| 147 |
+
# ÉTAPE 2: Extraction de l'audio 16k WAV à partir du fichier stabilisé.
|
| 148 |
+
extract_cmd = (
|
| 149 |
+
f'ffmpeg -hide_banner -loglevel error -y '
|
| 150 |
+
f'-i {shlex.quote(stabilized_mp4)} -vn -ac 1 -ar 16000 -f wav {shlex.quote(out_wav)}'
|
| 151 |
+
)
|
| 152 |
+
print("RUN: Extraction de l'audio depuis le fichier stabilisé...")
|
| 153 |
+
run_cmd(extract_cmd)
|
| 154 |
+
|
| 155 |
+
# Nettoyage du fichier intermédiaire stabilisé
|
| 156 |
+
if os.path.exists(stabilized_mp4):
|
| 157 |
+
os.remove(stabilized_mp4)
|
| 158 |
+
|
| 159 |
def clean_audio(wav_path, target_sr=16000):
|
| 160 |
+
"""Load audio, ensure mono, resample to target_sr, normalize, write cleaned wav."""
|
| 161 |
audio, sr = sf.read(wav_path)
|
| 162 |
+
if audio.ndim == 2:
|
| 163 |
+
audio = audio.mean(axis=1)
|
| 164 |
if sr != target_sr:
|
| 165 |
audio = librosa.resample(audio.astype(float), orig_sr=sr, target_sr=target_sr)
|
| 166 |
sr = target_sr
|
| 167 |
+
max_val = np.max(np.abs(audio)) if audio.size > 0 else 0.0
|
|
|
|
|
|
|
|
|
|
| 168 |
if max_val > 1e-6:
|
| 169 |
+
audio = audio / max_val * 0.9
|
| 170 |
clean_path = str(Path(wav_path).with_name(Path(wav_path).stem + "_clean.wav"))
|
| 171 |
sf.write(clean_path, audio, sr)
|
| 172 |
return clean_path, audio, sr
|
| 173 |
+
|
| 174 |
+
# ---------------------------- # TRANSCRIPTION, ETC. (Inchangé) # ----------------------------
|
| 175 |
+
# Les autres fonctions (transcribe, keep_bambara, pack, align_heuristic, etc.)
|
| 176 |
+
# restent les mêmes que dans la version V4.7.
|
| 177 |
|
|
|
|
|
|
|
|
|
|
| 178 |
def transcribe(model, wav_path):
|
| 179 |
+
if not hasattr(model, "transcribe"):
|
| 180 |
+
raise RuntimeError("Le modèle ne supporte pas model.transcribe()")
|
| 181 |
out = model.transcribe([wav_path])
|
| 182 |
+
if isinstance(out, list):
|
| 183 |
+
if len(out) == 0:
|
| 184 |
+
return ""
|
| 185 |
+
first = out[0]
|
| 186 |
+
if isinstance(first, str):
|
| 187 |
+
return first.strip()
|
| 188 |
+
if hasattr(first, "text"):
|
| 189 |
+
return first.text.strip()
|
| 190 |
+
return str(first).strip()
|
| 191 |
+
if hasattr(out, "text"):
|
| 192 |
+
return out.text.strip()
|
| 193 |
return str(out).strip()
|
| 194 |
+
|
| 195 |
+
def keep_bambara(words):
|
| 196 |
+
res = []
|
| 197 |
+
for w in words:
|
| 198 |
+
wl = w.lower()
|
| 199 |
+
if any(c in wl for c in ["ɛ","ɔ","ŋ"]) or sum(1 for c in wl if c in "aeiou") >= 2:
|
| 200 |
+
res.append(w)
|
| 201 |
+
return res
|
| 202 |
+
|
| 203 |
+
MAX_CHARS = 45; MIN_DUR = 0.3; MAX_DUR = 3.2; MAX_WORDS = 8
|
| 204 |
+
|
| 205 |
+
def wrap2(txt):
|
| 206 |
+
parts = textwrap.wrap(txt, MAX_CHARS)
|
| 207 |
+
if len(parts) <= 1:
|
| 208 |
+
return txt
|
| 209 |
+
mid = len(txt) // 2
|
| 210 |
+
left = txt.rfind(" ", 0, mid)
|
| 211 |
+
right = txt.find(" ", mid)
|
| 212 |
+
cut = left if (mid - left) <= ((right - mid) if right != -1 else 1e9) else right
|
| 213 |
+
l1 = txt[:cut].strip(); l2 = txt[cut:].strip()
|
| 214 |
+
return l1 + "\n" + l2 if l2 else l1
|
| 215 |
+
|
| 216 |
+
def pack(spans, total):
|
| 217 |
+
tmp = []
|
| 218 |
+
for s, e, t in spans:
|
| 219 |
+
s = max(0, min(s, total)); e = max(0, min(e, total))
|
| 220 |
+
if e <= s or not t.strip(): continue
|
| 221 |
+
tmp.append((s, e, t.strip()))
|
| 222 |
+
merged = []
|
| 223 |
+
for seg in tmp:
|
| 224 |
+
if not merged:
|
| 225 |
+
merged.append(seg); continue
|
| 226 |
+
ps, pe, pt = merged[-1]; s, e, t = seg
|
| 227 |
+
if (e - s) < MIN_DUR or (s - pe) < 0.1:
|
| 228 |
+
merged[-1] = (ps, max(pe, e), (pt + " " + t).strip())
|
| 229 |
+
else:
|
| 230 |
+
merged.append(seg)
|
| 231 |
+
out = []; last_end = 0
|
| 232 |
+
for s, e, t in merged:
|
| 233 |
+
dur = e - s; words = t.split()
|
| 234 |
+
blocks = [" ".join(words[i:i+MAX_WORDS]) for i in range(0, len(words), MAX_WORDS)]
|
| 235 |
+
step = dur / max(1, len(blocks))
|
| 236 |
+
base = s
|
| 237 |
+
for b in blocks:
|
| 238 |
+
st = base; en = min(base + step, e); base = en
|
| 239 |
+
if en <= st: en = min(st + 0.05, total)
|
| 240 |
+
txt = wrap2(b)
|
| 241 |
+
if st < last_end:
|
| 242 |
+
st = last_end + 1e-3; en = max(en, st + 0.05)
|
| 243 |
+
out.append((st, en, txt)); last_end = en
|
| 244 |
+
return out
|
| 245 |
+
|
| 246 |
+
def align_heuristic(words, total_dur):
|
| 247 |
+
total = total_dur
|
| 248 |
+
if not words:
|
| 249 |
+
return pack([], total)
|
| 250 |
+
|
| 251 |
+
spans = []
|
| 252 |
+
blocks = [" ".join(words[i:i+MAX_WORDS]) for i in range(0, len(words), MAX_WORDS)]
|
| 253 |
+
num_blocks = len(blocks)
|
| 254 |
+
|
| 255 |
+
max_step = min(MAX_DUR, total / num_blocks if num_blocks > 0 else total)
|
| 256 |
+
|
| 257 |
+
base = 0.0
|
| 258 |
+
for block in blocks:
|
| 259 |
+
st = base; en = min(base + max_step, total)
|
| 260 |
+
spans.append((st, en, block))
|
| 261 |
+
base = en
|
| 262 |
+
|
| 263 |
+
return pack(spans, total)
|
| 264 |
+
|
| 265 |
|
| 266 |
+
def segment_and_align(model, audio, sr, total_dur, mode):
|
| 267 |
+
"""Découpe l'audio, tente alignement CTC Segmentation, fallback Heuristique."""
|
| 268 |
+
segment_samples = int(SEGMENT_DURATION * sr)
|
| 269 |
+
total_samples = len(audio)
|
| 270 |
+
all_subs = []
|
| 271 |
+
|
| 272 |
+
for i in range(0, total_samples, segment_samples):
|
| 273 |
+
start_sample = i
|
| 274 |
+
end_sample = min(i + segment_samples, total_samples)
|
| 275 |
+
time_offset = start_sample / sr
|
| 276 |
+
|
| 277 |
+
segment_audio = audio[start_sample:end_sample]
|
| 278 |
+
segment_duration = (end_sample - start_sample) / sr
|
| 279 |
+
|
| 280 |
+
tmp_fd, tmp_seg_wav = tempfile.mkstemp(suffix=f"_seg_{i}.wav")
|
| 281 |
+
os.close(tmp_fd)
|
| 282 |
+
sf.write(tmp_seg_wav, segment_audio, sr)
|
| 283 |
+
|
| 284 |
+
try:
|
| 285 |
+
segment_text = transcribe(model, tmp_seg_wav)
|
| 286 |
+
words = keep_bambara(segment_text.split())
|
| 287 |
|
| 288 |
+
subs = None
|
| 289 |
+
if HAS_CTC_SEGMENTATION and words and mode in ["rnnt", "ctc"]:
|
| 290 |
+
try:
|
| 291 |
+
x = torch.tensor(segment_audio).float().unsqueeze(0).to(DEVICE)
|
| 292 |
+
ln = torch.tensor([x.shape[1]]).to(DEVICE)
|
| 293 |
+
|
| 294 |
+
with torch.no_grad():
|
| 295 |
+
logits, _ = model.forward(input_signal=x, input_signal_length=ln)
|
| 296 |
+
if isinstance(logits, tuple):
|
| 297 |
+
logits = logits[0]
|
| 298 |
+
|
| 299 |
+
time_per_frame = segment_duration / max(1, logits.shape[1])
|
| 300 |
+
|
| 301 |
+
try:
|
| 302 |
+
raw = model.tokenizer.vocab
|
| 303 |
+
vocab = list(raw.keys()) if isinstance(raw, dict) else list(raw)
|
| 304 |
+
except Exception:
|
| 305 |
+
vocab = None
|
| 306 |
+
|
| 307 |
+
cfg = CtcSegmentationParameters()
|
| 308 |
+
if vocab:
|
| 309 |
+
cfg.char_list = vocab
|
| 310 |
+
|
| 311 |
+
gt = prepare_text(cfg, words)[0]
|
| 312 |
+
|
| 313 |
+
# CORRECTION DU DÉBALLAGE (STAR-UNPACKING)
|
| 314 |
+
timing, *others = ctc_segmentation(cfg, logits.detach().cpu().numpy()[0], gt)
|
| 315 |
+
|
| 316 |
+
spans = []
|
| 317 |
+
for k in range(len(words)):
|
| 318 |
+
start_time = timing[k] * time_per_frame
|
| 319 |
+
end_time = timing[k+1] * time_per_frame if k + 1 < len(timing) else segment_duration
|
| 320 |
+
spans.append((start_time, end_time, words[k]))
|
| 321 |
|
| 322 |
+
subs = pack(spans, segment_duration)
|
| 323 |
+
|
| 324 |
+
except Exception as e:
|
| 325 |
+
print(f"[WARN] CTC Segmentation échoué pour le segment à {time_offset:.2f}s ({e}) -> Fallback Heuristique")
|
| 326 |
+
subs = align_heuristic(words, segment_duration)
|
| 327 |
+
else:
|
| 328 |
+
subs = align_heuristic(words, segment_duration)
|
| 329 |
|
| 330 |
+
if subs:
|
| 331 |
+
for start, end, text in subs:
|
| 332 |
+
all_subs.append((start + time_offset, end + time_offset, text))
|
| 333 |
+
|
| 334 |
+
except Exception as e:
|
| 335 |
+
print(f"Échec critique de la transcription/alignement du segment à {time_offset:.2f}s: {e}")
|
| 336 |
+
|
| 337 |
+
finally:
|
| 338 |
+
if os.path.exists(tmp_seg_wav):
|
| 339 |
+
os.remove(tmp_seg_wav)
|
| 340 |
|
| 341 |
+
return pack(all_subs, total_dur)
|
| 342 |
+
|
| 343 |
+
def burn(video_path, subs, output_path=None):
|
| 344 |
+
if output_path is None:
|
| 345 |
+
output_path = "RobotsMali_Subtitled.mp4"
|
| 346 |
+
|
| 347 |
+
tmp_fd, tmp_srt = tempfile.mkstemp(suffix=".srt")
|
| 348 |
+
os.close(tmp_fd)
|
| 349 |
+
|
| 350 |
+
def sec_to_srt(t):
|
| 351 |
+
h = int(t // 3600); m = int((t % 3600) // 60); s = int(t % 60); ms = int((t - int(t)) * 1000)
|
| 352 |
+
return f"{h:02}:{m:02}:{s:02},{ms:03}"
|
| 353 |
+
|
| 354 |
+
with open(tmp_srt, "w", encoding="utf-8") as f:
|
| 355 |
+
for i, (start, end, text) in enumerate(subs, 1):
|
| 356 |
+
f.write(f"{i}\n{sec_to_srt(start)} --> {sec_to_srt(end)}\n{text}\n\n")
|
| 357 |
+
|
| 358 |
+
vf = f"subtitles={shlex.quote(tmp_srt)}:force_style='Fontsize=22,PrimaryColour=&HFFFFFF&,OutlineColour=&H000000&'"
|
| 359 |
+
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)}'
|
| 360 |
+
|
| 361 |
+
try:
|
| 362 |
+
run_cmd(cmd)
|
| 363 |
+
finally:
|
| 364 |
+
if os.path.exists(tmp_srt):
|
| 365 |
+
os.remove(tmp_srt)
|
| 366 |
+
|
| 367 |
+
return output_path
|
| 368 |
+
|
| 369 |
+
# ---------------------------- # PIPELINE PRINCIPAL (V4.8) # ----------------------------
|
| 370 |
+
def pipeline(video_input, model_name):
|
| 371 |
+
try:
|
| 372 |
+
if isinstance(video_input, dict) and "tmp_path" in video_input:
|
| 373 |
+
video_path = video_input["tmp_path"]
|
| 374 |
+
else:
|
| 375 |
+
video_path = video_input
|
| 376 |
+
|
| 377 |
+
# Tente d'obtenir la durée via ffprobe (pour un contrôle rapide)
|
| 378 |
+
duration = ffprobe_duration(video_path)
|
| 379 |
+
|
| 380 |
+
tmp_fd, tmp_wav = tempfile.mkstemp(suffix=".wav")
|
| 381 |
+
os.close(tmp_fd)
|
| 382 |
+
|
| 383 |
+
# Extraction audio robuste (tentative de réparation/remuxage via ffmpeg)
|
| 384 |
+
extract_audio(video_path, tmp_wav)
|
| 385 |
+
clean_wav, audio, sr = clean_audio(tmp_wav)
|
| 386 |
+
|
| 387 |
+
# LOGIQUE DE FALLBACK : Si ffprobe a échoué, calcule la durée à partir du fichier WAV extrait
|
| 388 |
+
if duration is None:
|
| 389 |
+
if len(audio) > 0:
|
| 390 |
+
duration = len(audio) / sr
|
| 391 |
+
print(f"[WARN] FFprobe échoué. Durée recalculée à partir de l'audio extrait : {duration:.2f}s")
|
| 392 |
+
else:
|
| 393 |
+
raise RuntimeError("Impossible d'obtenir une durée non nulle de la vidéo, même après extraction audio robuste.")
|
| 394 |
+
|
| 395 |
+
model = load_model(model_name)
|
| 396 |
+
mode = MODELS[model_name][1]
|
| 397 |
+
|
| 398 |
+
subs = segment_and_align(model, audio, sr, duration, mode)
|
| 399 |
+
|
| 400 |
+
if not subs:
|
| 401 |
+
return ("⚠️ Aucun sous-titre utilisable (sub list vide)", None)
|
| 402 |
+
|
| 403 |
+
out_video = burn(video_path, subs)
|
| 404 |
+
return ("✅ Terminé avec succès", out_video)
|
| 405 |
+
|
| 406 |
+
except Exception as e:
|
| 407 |
+
traceback.print_exc()
|
| 408 |
+
return (f"❌ Erreur — {str(e)}", None)
|
| 409 |
|
| 410 |
+
|
| 411 |
+
# ---------------------------- # INTERFACE GRADIO # ----------------------------
|
|
|
|
| 412 |
with gr.Blocks(title="RobotsMali - Sous-titrage") as demo:
|
| 413 |
+
gr.Markdown("## 🤖 RobotsMali — Sous-titrage Bambara (Colab Ready - Audio Max Robuste)")
|
| 414 |
+
gr.Markdown("L'extraction audio est maintenant ultra-robuste. Si vous utilisez la webcam ou un fichier téléchargé, ce script devrait pouvoir le traiter.")
|
| 415 |
+
|
| 416 |
+
# Composant Video sans 'examples'
|
| 417 |
+
v = gr.Video(label="Vidéo à sous-titrer (Fichier ou Webcam)")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 418 |
|
| 419 |
+
# Utilisation de gr.Examples séparé pour la compatibilité
|
| 420 |
+
gr.Examples(
|
| 421 |
+
examples=VIDEO_EXAMPLES,
|
| 422 |
+
inputs=v,
|
| 423 |
+
label="Exemples de vidéos à tester (Téléchargez d'abord le fichier dans Colab pour utiliser ce chemin)"
|
| 424 |
+
)
|
| 425 |
+
|
| 426 |
+
m = gr.Dropdown(list(MODELS.keys()), value="Soloba V1 (CTC)", label="Modèle ASR Bambara")
|
| 427 |
+
b = gr.Button("▶️ Générer les Sous-titres Incrustés")
|
| 428 |
+
s = gr.Markdown(label="Statut")
|
| 429 |
+
o = gr.Video(label="Vidéo sous-titrée (Format MP4 H.264)")
|
| 430 |
+
|
| 431 |
b.click(pipeline, [v, m], [s, o])
|
| 432 |
+
|
| 433 |
+
demo.launch(share=True, debug=True)
|
|
|