diff --git a/.env.example b/.env.example
index 4f98eb8..06ad168 100644
--- a/.env.example
+++ b/.env.example
@@ -12,6 +12,7 @@ LOCAL_AI_CACHE_DIR=.local-ai/cache
 # Windows면 `python` 또는 `py -3`
 LOCAL_AI_PYTHON=
 LOCAL_STT_MODEL=tiny
+# CUDA dll 오류가 나면 `cpu`
 LOCAL_STT_DEVICE=auto
 LOCAL_STT_COMPUTE_TYPE=auto
 LOCAL_STT_BEAM_SIZE=1
diff --git a/README.md b/README.md
index 776d6b0..2b604e2 100644
--- a/README.md
+++ b/README.md
@@ -118,6 +118,14 @@ Discord 모드에서만 필수:
 - `DEBUG_TEXT_EVENTS`
   - `true`면 transcript/reply를 콘솔에 같이 출력
 
+Windows에서 STT 시작 시 `cublas64_12.dll` 같은 CUDA 오류가 나면:
+
+```env
+LOCAL_STT_DEVICE=cpu
+```
+
+로 두면 바로 우회됩니다. 최신 코드는 `auto`일 때도 가능한 경우 CPU로 자동 fallback 합니다.
+
 ## 속도 우선 기본값
 
 - STT 기본 모델은 `tiny`
diff --git a/python/local_stt_worker.py b/python/local_stt_worker.py
index a94783f..fd910d3 100644
--- a/python/local_stt_worker.py
+++ b/python/local_stt_worker.py
@@ -58,14 +58,36 @@ class SttWorker:
         from faster_whisper import WhisperModel
 
         self.model_name = os.environ.get("LOCAL_STT_MODEL", "tiny").strip() or "tiny"
-        self.device = resolve_device()
-        self.compute_type = resolve_compute_type(self.device)
+        requested_device = resolve_device()
+        requested_compute_type = resolve_compute_type(requested_device)
         self.beam_size = int(os.environ.get("LOCAL_STT_BEAM_SIZE", "1"))
-        self.model = WhisperModel(
-            self.model_name,
-            device=self.device,
-            compute_type=self.compute_type,
-        )
+        auto_requested = os.environ.get("LOCAL_STT_DEVICE", "auto").strip().lower() in {"", "auto"}
+
+        try:
+            self.model = WhisperModel(
+                self.model_name,
+                device=requested_device,
+                compute_type=requested_compute_type,
+            )
+            self.device = requested_device
+            self.compute_type = requested_compute_type
+        except RuntimeError as exc:
+            lowered = str(exc).lower()
+            should_fallback = auto_requested and requested_device == "cuda" and any(
+                token in lowered for token in ("cublas", "cudnn", "cuda")
+            )
+            if not should_fallback:
+                raise
+
+            log("CUDA runtime is incomplete; falling back to CPU STT")
+            self.model = WhisperModel(
+                self.model_name,
+                device="cpu",
+                compute_type=resolve_compute_type("cpu"),
+            )
+            self.device = "cpu"
+            self.compute_type = resolve_compute_type("cpu")
+
         log(
             f"local-stt ready model={self.model_name} device={self.device} compute={self.compute_type} beam={self.beam_size}"
         )