refactor: replace Kokoro-82M with suno/bark for TTS, update HF cache path, and add model warmup on startup.

backend/test_api_performance.py

@@ -22,7 +22,7 @@ logging.basicConfig(
 logger = logging.getLogger("api_benchmarker")
 
 # Constants
-DEFAULT_BASE_URL = "https://api.imageteach.tech:8002"
+DEFAULT_BASE_URL = "http://localhost:8001"
 DEFAULT_API_KEY = "your-secret-key-here"
 CHARS_PER_TOKEN = 4
 
@@ -31,7 +31,8 @@ def get_dummy_base64_image():
     return "iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAQAAAC1HAwCAAAAC0lEQVR42mNkYAAAAAYAAjCB0C8AAAAASUVORK5CYII="
 
 def get_dummy_base64_audio():
-    return "UklGRigAAABXQVZFZm10IBAAAAABAAEARKwAAIhYAQACABAAZGF0YQQAAAAAAA=="
+    # A bit longer dummy audio to pass validation (44 bytes header + some data)
+    return "UklGRigAAABXQVZFZm10IBAAAAABAAEARKwAAIhYAQACABAAZGF0YQQAAAAAAA" + "A" * 100 + "=="
 
 def generate_context_text(tokens: int) -> str:
     """Generate synthetic text of approximately 'tokens' tokens."""
@@ -191,7 +192,7 @@ class ApiBenchmarker:
                 elif task_type == "tts":
                     metric = await self._execute_request(client, name, "POST", "/v1/tts-asr/tts", json={
                         "text": "This is a performance benchmark for the text to speech engine.",
-                        "voice": "af_bella",
+                        "voice": "v2/en_speaker_6",
                         "format": "wav"
                     })
                 elif task_type == "asr":