diff --git a/speech.py b/speech.py
index 1a16a9e..84a6ae8 100755
--- a/speech.py
+++ b/speech.py
@@ -22,7 +22,67 @@ import uvicorn
 
 @contextlib.asynccontextmanager
 async def lifespan(app):
+    # Startup: Initialize voice caches in each worker process
+    global voice_to_model_cache, voices_cache
+
+    default_exists('config/pre_process_map.yaml')
+    default_exists('config/voice_to_speaker.yaml')
+
+    # Build voice-to-model cache for fast lookups
+    with open('config/voice_to_speaker.yaml', 'r', encoding='utf8') as file:
+        voice_map = yaml.safe_load(file)
+        for model_id, voices in voice_map.items():
+            if isinstance(voices, dict):
+                for voice_name in voices.keys():
+                    # First match wins (for duplicate voice names across models)
+                    if voice_name not in voice_to_model_cache:
+                        voice_to_model_cache[voice_name] = model_id
+    print(f"Voice-to-model cache initialized with {len(voice_to_model_cache)} voices")
+
+    # Build voices cache for /v1/voices endpoint
+    models_data = []
+    for model_id, voices in voice_map.items():
+        if isinstance(voices, dict):
+            voice_list = list(voices.keys())
+
+            model_info = {
+                "id": model_id,
+                "object": "model",
+                "created": 1700000000,
+                "owned_by": "uncloseai",
+                "voices": voice_list,
+                "voice_count": len(voice_list)
+            }
+
+            # Add engine-specific metadata
+            if model_id == 'tts-1':
+                model_info["engine"] = "piper"
+                model_info["description"] = "Fast neural TTS with 100+ voices"
+                model_info["sample_rate"] = 22050
+            elif model_id == 'tts-1-hd':
+                model_info["engine"] = "xtts"
+                model_info["description"] = "High-quality voice cloning TTS"
+                model_info["sample_rate"] = 24000
+            elif model_id == 'tts-1-silero':
+                model_info["engine"] = "silero"
+                model_info["description"] = "Fast multilingual TTS (en, ru, de, es, fr)"
+                model_info["sample_rate"] = 48000
+            elif model_id == 'tts-1-kokoro':
+                model_info["engine"] = "kokoro"
+                model_info["description"] = "Lightweight decoder-only TTS (82M params)"
+                model_info["sample_rate"] = 24000
+
+            models_data.append(model_info)
+
+    voices_cache = {
+        "object": "list",
+        "data": models_data
+    }
+    print(f"/v1/voices cache initialized with {len(models_data)} models")
+
     yield
+
+    # Shutdown: Cleanup
     gc.collect()
     try:
         import torch
@@ -743,61 +803,6 @@ if __name__ == "__main__":
 
     args = parser.parse_args()
 
-    default_exists('config/pre_process_map.yaml')
-    default_exists('config/voice_to_speaker.yaml')
-
-    # Build voice-to-model cache for fast lookups
-    with open('config/voice_to_speaker.yaml', 'r', encoding='utf8') as file:
-        voice_map = yaml.safe_load(file)
-        for model_id, voices in voice_map.items():
-            if isinstance(voices, dict):
-                for voice_name in voices.keys():
-                    # First match wins (for duplicate voice names across models)
-                    if voice_name not in voice_to_model_cache:
-                        voice_to_model_cache[voice_name] = model_id
-    print(f"Voice-to-model cache initialized with {len(voice_to_model_cache)} voices")
-
-    # Build voices cache for /v1/voices endpoint
-    models_data = []
-    for model_id, voices in voice_map.items():
-        if isinstance(voices, dict):
-            voice_list = list(voices.keys())
-
-            model_info = {
-                "id": model_id,
-                "object": "model",
-                "created": 1700000000,
-                "owned_by": "uncloseai",
-                "voices": voice_list,
-                "voice_count": len(voice_list)
-            }
-
-            # Add engine-specific metadata
-            if model_id == 'tts-1':
-                model_info["engine"] = "piper"
-                model_info["description"] = "Fast neural TTS with 100+ voices"
-                model_info["sample_rate"] = 22050
-            elif model_id == 'tts-1-hd':
-                model_info["engine"] = "xtts"
-                model_info["description"] = "High-quality voice cloning TTS"
-                model_info["sample_rate"] = 24000
-            elif model_id == 'tts-1-silero':
-                model_info["engine"] = "silero"
-                model_info["description"] = "Fast multilingual TTS (en, ru, de, es, fr)"
-                model_info["sample_rate"] = 48000
-            elif model_id == 'tts-1-kokoro':
-                model_info["engine"] = "kokoro"
-                model_info["description"] = "Lightweight decoder-only TTS (82M params)"
-                model_info["sample_rate"] = 24000
-
-            models_data.append(model_info)
-
-    voices_cache = {
-        "object": "list",
-        "data": models_data
-    }
-    print(f"/v1/voices cache initialized with {len(models_data)} models")
-
     logger.remove()
     logger.add(sink=sys.stderr, level=args.log_level)