modelpy

.env

@@ -1,7 +1,7 @@
-SAMPLE_RATE = 16000
-CHANNELS = 1
-SAMPLE_WIDTH = 2
+SAMPLE_RATE=16000
+CHANNELS=1
+SAMPLE_WIDTH=2
 
-CHUNK_DURATION_MS = 500  
-DB_PATH = "audio_sessions.db"
-AUDIO_STORAGE_DIR = "audio_chunks"
+CHUNK_DURATION_MS=500
+DB_PATH=/tmp/audio_sessions.db
+AUDIO_STORAGE_DIR=/tmp/audio_chunks

src/database.py

@@ -17,7 +17,9 @@ def init_db():
             actual_duration_ms REAL,
             sura_number INTEGER,
             ayat_begin INTEGER,
-            ayat_end INTEGER
+            ayat_end INTEGER,
+            word_begin INTEGER,
+            word_end INTEGER
         )
     ''')
     conn.commit()

src/model.py

@@ -1,58 +1,42 @@
 import os
 from transformers import WhisperProcessor, WhisperForConditionalGeneration
 
-# Hugging Face model repo
 MODEL_NAME = "tarteel-ai/whisper-tiny-ar-quran"
 
-# Local cache path (works locally, or will fallback to /tmp in Spaces)
-LOCAL_SAVE_DIR = os.path.abspath("./whisper-tiny-ar-quran-local")
-HF_SPACE_CACHE = "/tmp/whisper-tiny-ar-quran"
+# Use /tmp for all writable storage
+CACHE_DIR = "/tmp/hf_cache"
+LOCAL_SAVE_DIR = "/tmp/whisper-tiny-ar-quran-local"
+
+# Make sure dirs exist
+os.makedirs(CACHE_DIR, exist_ok=True)
+os.makedirs(LOCAL_SAVE_DIR, exist_ok=True)
+
+os.environ["TRANSFORMERS_CACHE"] = CACHE_DIR
+os.environ["HF_HOME"] = CACHE_DIR
+os.environ["HF_HUB_CACHE"] = CACHE_DIR
 
-# Globals (singleton pattern)
 _model = None
 _processor = None
 
-
 def load_model():
-    """
-    Load Whisper model + processor.
-    Priority:
-    1. Reuse in-memory model if already loaded.
-    2. Load from local_save_dir if exists.
-    3. Otherwise, download from HuggingFace Hub.
-       - If running on HuggingFace Spaces, use /tmp to avoid permission issues.
-       - Save locally when possible.
-    """
     global _model, _processor
 
     if _model is not None and _processor is not None:
         return _model, _processor
 
-    # Pick best cache dir (local if writable, else /tmp)
-    target_dir = LOCAL_SAVE_DIR
-    try:
-        os.makedirs(LOCAL_SAVE_DIR, exist_ok=True)
-        test_path = os.path.join(LOCAL_SAVE_DIR, "write_test.txt")
-        with open(test_path, "w") as f:
-            f.write("ok")
-        os.remove(test_path)
-    except (PermissionError, OSError):
-        target_dir = HF_SPACE_CACHE
-        os.makedirs(target_dir, exist_ok=True)
-
-    local_config_file = os.path.join(target_dir, "preprocessor_config.json")
+    local_config_file = os.path.join(LOCAL_SAVE_DIR, "preprocessor_config.json")
 
     if os.path.exists(local_config_file):
-        _processor = WhisperProcessor.from_pretrained(target_dir)
-        _model = WhisperForConditionalGeneration.from_pretrained(target_dir)
+        _processor = WhisperProcessor.from_pretrained(LOCAL_SAVE_DIR)
+        _model = WhisperForConditionalGeneration.from_pretrained(LOCAL_SAVE_DIR)
     else:
-        _processor = WhisperProcessor.from_pretrained(MODEL_NAME)
-        _model = WhisperForConditionalGeneration.from_pretrained(MODEL_NAME)
+        _processor = WhisperProcessor.from_pretrained(MODEL_NAME, cache_dir=CACHE_DIR)
+        _model = WhisperForConditionalGeneration.from_pretrained(MODEL_NAME, cache_dir=CACHE_DIR)
 
-        _model.save_pretrained(target_dir)
-        _processor.save_pretrained(target_dir)
+        _model.save_pretrained(LOCAL_SAVE_DIR)
+        _processor.save_pretrained(LOCAL_SAVE_DIR)
 
-    # Patch configs (for Arabic transcribe)
+    # Patch configs
     if not hasattr(_model.generation_config, "lang_to_id"):
         _model.generation_config.lang_to_id = {"arabic": 50361}
     if not hasattr(_model.generation_config, "task_to_id"):
@@ -61,5 +45,5 @@ def load_model():
     _model.generation_config.language = "arabic"
     _model.generation_config.task = "transcribe"
 
-    print(f"✅ Model and processor loaded from: {target_dir}")
-    return _model, _processor
+    print(f"✅ Model and processor loaded from: {LOCAL_SAVE_DIR}")
+    return _model, _processor