main: background self-warm thread keeps the model hot 24/7

User is on a paid always-on (never-sleep) Space, yet the first request after a
day idle still took ~30s. Cause: the model path goes cold during idle (memory
paged out / CPU throttled) and /warmup never runs the model, so it can't keep it
warm. Added a daemon thread that runs a tiny LOCAL Wav2Vec2 inference every
SELF_WARM_INTERVAL_SEC (default 240s) — forward pass only, NO STT API calls
(those cost money) — so the model never goes cold regardless of app activity.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

main.py

@@ -13,6 +13,7 @@ import sys
 import uuid
 import time
 import tempfile
+import threading
 from contextlib import asynccontextmanager
 
 from fastapi import FastAPI, File, Form, UploadFile, HTTPException, Header
@@ -80,6 +81,14 @@ VALID_WORDS = {
 }
 
 
+# How often the background thread re-runs a tiny LOCAL inference to keep the
+# Wav2Vec2 model hot. The server is always-on, but after a day idle the model
+# path goes cold (memory paged out / CPU throttled) and the first real request
+# paid ~30s. /warmup never runs the model, so it can't prevent this. 240s keeps
+# the model warm with negligible cost. Override via env if needed.
+SELF_WARM_INTERVAL_SEC = int(os.environ.get("SELF_WARM_INTERVAL_SEC", "240"))
+
+
 @asynccontextmanager
 async def lifespan(app: FastAPI):
     """Pre-load Wav2Vec2 model AND Supabase client before accepting requests.
@@ -122,6 +131,34 @@ async def lifespan(app: FastAPI):
     except Exception as e:
         print(f"[WARMUP] dummy scoring setup failed: {e}")
 
+    # Background self-warm: keep the local Wav2Vec2 model HOT 24/7. The server is
+    # always-on (paid, never-sleep), but after a day idle the model path goes cold
+    # and the first real request paid ~30s. This daemon thread runs a tiny LOCAL
+    # inference every few minutes — Wav2Vec2 forward pass only, NO STT API calls
+    # (those cost money) — so the model never goes cold, independent of the app.
+    def _self_warm_loop():
+        import time as _time
+        import numpy as np
+        import soundfile as sf
+        warm_wav = os.path.join(tempfile.gettempdir(), "selfwarm.wav")
+        try:
+            rng = np.random.default_rng(1)
+            sf.write(warm_wav, (rng.standard_normal(8000) * 0.01).astype("float32"), 16000)
+        except Exception as e:
+            print(f"[SELF-WARM] setup failed, warmer not started: {e}")
+            return
+        from phoneme_extractor import shin_vs_samekh
+        while True:
+            _time.sleep(SELF_WARM_INTERVAL_SEC)
+            try:
+                shin_vs_samekh(warm_wav)  # local Wav2Vec2 forward pass, no network
+                print("[SELF-WARM] model kept hot")
+            except Exception as e:
+                print(f"[SELF-WARM] tick failed: {e}")
+
+    threading.Thread(target=_self_warm_loop, daemon=True).start()
+    print(f"[SELF-WARM] background warmer started (every {SELF_WARM_INTERVAL_SEC}s)")
+
     yield