Spaces:
Runtime error
Runtime error
Update app.py
Browse files
app.py
CHANGED
|
@@ -8,16 +8,15 @@ import gradio as gr
|
|
| 8 |
from huggingface_hub import snapshot_download
|
| 9 |
from nemo.collections import asr as nemo_asr
|
| 10 |
|
| 11 |
-
# Configuration
|
| 12 |
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
|
| 13 |
SEGMENT_DURATION = 10.0
|
| 14 |
|
| 15 |
-
#
|
| 16 |
MODELS = {
|
| 17 |
"Soloni V3 (Recommandé CPU)": "RobotsMali/soloni-114m-tdt-ctc-v3"
|
| 18 |
}
|
| 19 |
|
| 20 |
-
# Cache pour éviter de recharger le modèle à chaque clic
|
| 21 |
_cache = {}
|
| 22 |
|
| 23 |
# ==========================
|
|
@@ -28,12 +27,12 @@ def get_model(model_name):
|
|
| 28 |
return _cache[model_name]
|
| 29 |
|
| 30 |
repo_id = MODELS[model_name]
|
| 31 |
-
print(f"⏳
|
| 32 |
|
| 33 |
-
# Téléchargement
|
| 34 |
folder = snapshot_download(repo_id, local_dir_use_symlinks=False)
|
| 35 |
|
| 36 |
-
#
|
| 37 |
try:
|
| 38 |
nemo_file = next(
|
| 39 |
os.path.join(folder, f)
|
|
@@ -41,11 +40,9 @@ def get_model(model_name):
|
|
| 41 |
if f.endswith(".nemo")
|
| 42 |
)
|
| 43 |
except StopIteration:
|
| 44 |
-
raise FileNotFoundError("
|
| 45 |
|
| 46 |
-
|
| 47 |
-
|
| 48 |
-
# Utilisation de ASRModel pour l'auto-détection (CTC/RNNT)
|
| 49 |
model = nemo_asr.models.ASRModel.restore_from(
|
| 50 |
nemo_file,
|
| 51 |
map_location=torch.device(DEVICE)
|
|
@@ -56,100 +53,82 @@ def get_model(model_name):
|
|
| 56 |
return model
|
| 57 |
|
| 58 |
# ==========================
|
| 59 |
-
#
|
| 60 |
# ==========================
|
| 61 |
def pipeline(audio_path, model_name):
|
| 62 |
if not audio_path:
|
| 63 |
-
yield "❌
|
| 64 |
return
|
| 65 |
|
| 66 |
-
# Création d'un dossier temporaire unique pour les segments
|
| 67 |
tmp_dir = tempfile.mkdtemp()
|
| 68 |
|
| 69 |
try:
|
| 70 |
-
yield "⏳
|
| 71 |
|
| 72 |
-
#
|
| 73 |
-
#
|
| 74 |
-
|
| 75 |
f"ffmpeg -y -i '{audio_path}' -f segment -segment_time {SEGMENT_DURATION} "
|
| 76 |
f"-ac 1 -ar 16000 {tmp_dir}/seg_%03d.wav > /dev/null 2>&1"
|
| 77 |
)
|
| 78 |
-
os.system(command)
|
| 79 |
|
| 80 |
-
# Liste et tri des segments générés
|
| 81 |
segments = sorted(glob.glob(os.path.join(tmp_dir, "seg_*.wav")))
|
| 82 |
-
|
| 83 |
-
# Filtrer les fichiers trop petits (silences ou erreurs FFmpeg)
|
| 84 |
segments = [s for s in segments if os.path.getsize(s) > 1000]
|
| 85 |
|
| 86 |
if not segments:
|
| 87 |
-
yield "❌ Erreur
|
| 88 |
return
|
| 89 |
|
| 90 |
-
yield f"🎙️ Transcription
|
| 91 |
|
| 92 |
-
# Récupération du modèle
|
| 93 |
model = get_model(model_name)
|
| 94 |
|
| 95 |
-
# Inférence
|
| 96 |
with torch.inference_mode():
|
| 97 |
-
|
| 98 |
-
|
| 99 |
-
batch_size=4,
|
| 100 |
-
num_workers=0 # Important sur CPU pour éviter les fuites de mémoire
|
| 101 |
-
)
|
| 102 |
|
| 103 |
-
#
|
| 104 |
-
# Note: selon le modèle, results peut être une liste de chaînes ou un objet complexe
|
| 105 |
if isinstance(results, tuple):
|
| 106 |
-
|
| 107 |
else:
|
| 108 |
-
|
| 109 |
|
| 110 |
-
final_text = " ".join(
|
| 111 |
-
|
| 112 |
-
yield "✅ Transcription terminée", final_text
|
| 113 |
|
| 114 |
except Exception as e:
|
| 115 |
-
yield
|
| 116 |
|
| 117 |
finally:
|
| 118 |
-
# Nettoyage automatique du dossier temporaire
|
| 119 |
if os.path.exists(tmp_dir):
|
| 120 |
shutil.rmtree(tmp_dir)
|
| 121 |
|
| 122 |
# ==========================
|
| 123 |
-
# INTERFACE
|
| 124 |
# ==========================
|
| 125 |
-
with gr.Blocks(
|
| 126 |
-
gr.Markdown(""
|
| 127 |
-
|
| 128 |
-
Déposez un fichier audio pour obtenir une transcription automatique.
|
| 129 |
-
""")
|
| 130 |
|
| 131 |
with gr.Row():
|
| 132 |
with gr.Column():
|
| 133 |
-
audio_input = gr.Audio(type="filepath", label="
|
| 134 |
model_input = gr.Dropdown(
|
| 135 |
choices=list(MODELS.keys()),
|
| 136 |
value="Soloni V3 (Recommandé CPU)",
|
| 137 |
-
label="
|
| 138 |
)
|
| 139 |
-
btn = gr.Button("🚀
|
| 140 |
|
| 141 |
with gr.Column():
|
| 142 |
-
|
| 143 |
-
|
| 144 |
|
| 145 |
-
# Interaction
|
| 146 |
btn.click(
|
| 147 |
fn=pipeline,
|
| 148 |
inputs=[audio_input, model_input],
|
| 149 |
-
outputs=[
|
| 150 |
)
|
| 151 |
|
| 152 |
-
gr.Markdown("--- \n *Optimisé pour les environnements à ressources limitées.*")
|
| 153 |
-
|
| 154 |
if __name__ == "__main__":
|
| 155 |
demo.launch()
|
|
|
|
| 8 |
from huggingface_hub import snapshot_download
|
| 9 |
from nemo.collections import asr as nemo_asr
|
| 10 |
|
| 11 |
+
# Configuration
|
| 12 |
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
|
| 13 |
SEGMENT_DURATION = 10.0
|
| 14 |
|
| 15 |
+
# Modèle unique pour CPU (CTC est plus léger)
|
| 16 |
MODELS = {
|
| 17 |
"Soloni V3 (Recommandé CPU)": "RobotsMali/soloni-114m-tdt-ctc-v3"
|
| 18 |
}
|
| 19 |
|
|
|
|
| 20 |
_cache = {}
|
| 21 |
|
| 22 |
# ==========================
|
|
|
|
| 27 |
return _cache[model_name]
|
| 28 |
|
| 29 |
repo_id = MODELS[model_name]
|
| 30 |
+
print(f"⏳ Chargement du modèle : {repo_id}")
|
| 31 |
|
| 32 |
+
# Téléchargement local
|
| 33 |
folder = snapshot_download(repo_id, local_dir_use_symlinks=False)
|
| 34 |
|
| 35 |
+
# Trouve le fichier .nemo
|
| 36 |
try:
|
| 37 |
nemo_file = next(
|
| 38 |
os.path.join(folder, f)
|
|
|
|
| 40 |
if f.endswith(".nemo")
|
| 41 |
)
|
| 42 |
except StopIteration:
|
| 43 |
+
raise FileNotFoundError("Fichier .nemo introuvable.")
|
| 44 |
|
| 45 |
+
# Restauration générique (Auto-détecte CTC/RNNT)
|
|
|
|
|
|
|
| 46 |
model = nemo_asr.models.ASRModel.restore_from(
|
| 47 |
nemo_file,
|
| 48 |
map_location=torch.device(DEVICE)
|
|
|
|
| 53 |
return model
|
| 54 |
|
| 55 |
# ==========================
|
| 56 |
+
# LOGIQUE DE TRANSCRIPTION
|
| 57 |
# ==========================
|
| 58 |
def pipeline(audio_path, model_name):
|
| 59 |
if not audio_path:
|
| 60 |
+
yield "❌ Erreur", "Veuillez fournir un fichier audio."
|
| 61 |
return
|
| 62 |
|
|
|
|
| 63 |
tmp_dir = tempfile.mkdtemp()
|
| 64 |
|
| 65 |
try:
|
| 66 |
+
yield "⏳ Traitement audio...", ""
|
| 67 |
|
| 68 |
+
# Segmentation FFmpeg (16kHz Mono requis)
|
| 69 |
+
# On utilise une syntaxe simple pour éviter les erreurs de shell
|
| 70 |
+
os.system(
|
| 71 |
f"ffmpeg -y -i '{audio_path}' -f segment -segment_time {SEGMENT_DURATION} "
|
| 72 |
f"-ac 1 -ar 16000 {tmp_dir}/seg_%03d.wav > /dev/null 2>&1"
|
| 73 |
)
|
|
|
|
| 74 |
|
|
|
|
| 75 |
segments = sorted(glob.glob(os.path.join(tmp_dir, "seg_*.wav")))
|
|
|
|
|
|
|
| 76 |
segments = [s for s in segments if os.path.getsize(s) > 1000]
|
| 77 |
|
| 78 |
if not segments:
|
| 79 |
+
yield "❌ Erreur", "Impossible de segmenter l'audio."
|
| 80 |
return
|
| 81 |
|
| 82 |
+
yield f"🎙️ Transcription ({len(segments)} segments)...", ""
|
| 83 |
|
|
|
|
| 84 |
model = get_model(model_name)
|
| 85 |
|
|
|
|
| 86 |
with torch.inference_mode():
|
| 87 |
+
# Transcription par batch pour le CPU
|
| 88 |
+
results = model.transcribe(segments, batch_size=2, num_workers=0)
|
|
|
|
|
|
|
|
|
|
| 89 |
|
| 90 |
+
# Gestion des différents formats de sortie de NeMo
|
|
|
|
| 91 |
if isinstance(results, tuple):
|
| 92 |
+
text_list = results[0]
|
| 93 |
else:
|
| 94 |
+
text_list = results
|
| 95 |
|
| 96 |
+
final_text = " ".join(text_list)
|
| 97 |
+
yield "✅ Terminé", final_text
|
|
|
|
| 98 |
|
| 99 |
except Exception as e:
|
| 100 |
+
yield "❌ Erreur Système", str(e)
|
| 101 |
|
| 102 |
finally:
|
|
|
|
| 103 |
if os.path.exists(tmp_dir):
|
| 104 |
shutil.rmtree(tmp_dir)
|
| 105 |
|
| 106 |
# ==========================
|
| 107 |
+
# INTERFACE UTILISATEUR
|
| 108 |
# ==========================
|
| 109 |
+
with gr.Blocks() as demo:
|
| 110 |
+
gr.Markdown("# 🤖 RobotsMali ASR")
|
| 111 |
+
gr.Markdown("Outil de transcription automatique optimisé pour le CPU.")
|
|
|
|
|
|
|
| 112 |
|
| 113 |
with gr.Row():
|
| 114 |
with gr.Column():
|
| 115 |
+
audio_input = gr.Audio(type="filepath", label="Audio")
|
| 116 |
model_input = gr.Dropdown(
|
| 117 |
choices=list(MODELS.keys()),
|
| 118 |
value="Soloni V3 (Recommandé CPU)",
|
| 119 |
+
label="Modèle"
|
| 120 |
)
|
| 121 |
+
btn = gr.Button("🚀 Transcrire", variant="primary")
|
| 122 |
|
| 123 |
with gr.Column():
|
| 124 |
+
status_output = gr.Textbox(label="Statut", interactive=False)
|
| 125 |
+
text_output = gr.Textbox(label="Texte Transcrit", lines=10)
|
| 126 |
|
|
|
|
| 127 |
btn.click(
|
| 128 |
fn=pipeline,
|
| 129 |
inputs=[audio_input, model_input],
|
| 130 |
+
outputs=[status_output, text_output]
|
| 131 |
)
|
| 132 |
|
|
|
|
|
|
|
| 133 |
if __name__ == "__main__":
|
| 134 |
demo.launch()
|