Switch to Whisper medium model with lazy loading

Dockerfile

@@ -35,11 +35,11 @@ RUN python -c "from transformers import AutoTokenizer, AutoModelForSequenceClass
     AutoModelForSequenceClassification.from_pretrained('Cyberlace/swara-structure-model', cache_dir='/.cache'); \
     print('✅ Structure Model cached!')"
 
-# 2. Download Whisper Base Model (~140MB) - lighter and faster
-RUN python -c "import whisper; \
-    print('📥 Downloading Whisper base model...'); \
-    whisper.load_model('base', download_root='/.cache'); \
-    print('✅ Whisper base cached!')"
+# 2. Whisper will be downloaded on first use (lazy loading to avoid OOM during build)
+# RUN python -c "import whisper; \
+#     print('📥 Downloading Whisper medium model...'); \
+#     whisper.load_model('medium', download_root='/.cache'); \
+#     print('✅ Whisper medium cached!')"
 
 # 3. Download Sentence Transformer for Keywords (~420MB)
 RUN python -c "from sentence_transformers import SentenceTransformer; \

app/config.py

@@ -31,7 +31,7 @@ class Settings(BaseSettings):
     UPLOAD_DIR: str = os.getenv("UPLOAD_DIR", "./uploads")
     
     # Model Configuration
-    WHISPER_MODEL: str = os.getenv("WHISPER_MODEL", "base")
+    WHISPER_MODEL: str = os.getenv("WHISPER_MODEL", "medium")  # Changed to medium for better accuracy
     KATA_KUNCI_PATH: str = os.getenv("KATA_KUNCI_PATH", "./kata_kunci.json")
     
     # Device Configuration (CPU/GPU)