diff --git a/tests/entrypoints/openai_api/test_serving_speech.py b/tests/entrypoints/openai_api/test_serving_speech.py
index 573816a0e..53edd9dc0 100644
--- a/tests/entrypoints/openai_api/test_serving_speech.py
+++ b/tests/entrypoints/openai_api/test_serving_speech.py
@@ -12,6 +12,7 @@ import os
 import struct
 import wave
 from dataclasses import FrozenInstanceError, replace
+from http import HTTPStatus
 from inspect import Signature, signature
 from pathlib import Path
 from types import SimpleNamespace
@@ -1012,6 +1013,44 @@ class TestTTSMethods:
         assert speed == 1.25
         resolve_adapter.assert_not_called()
 
+    @pytest.mark.asyncio
+    async def test_create_speech_cuda_oom_returns_internal_server_error(
+        self,
+        speech_server,
+        mocker: MockerFixture,
+    ):
+        mocker.patch.object(speech_server, "_check_model", new=mocker.AsyncMock(return_value=None))
+        mocker.patch.object(
+            speech_server,
+            "_generate_audio_bytes",
+            new=mocker.AsyncMock(side_effect=torch.OutOfMemoryError("CUDA out of memory")),
+        )
+
+        response = await speech_server.create_speech(OpenAICreateSpeechRequest(input="Hello"))
+
+        assert isinstance(response, ErrorResponse)
+        assert response.error.code == HTTPStatus.INTERNAL_SERVER_ERROR
+        assert response.error.type == "InternalServerError"
+
+    @pytest.mark.asyncio
+    async def test_create_speech_unexpected_failure_returns_internal_server_error(
+        self,
+        speech_server,
+        mocker: MockerFixture,
+    ):
+        mocker.patch.object(speech_server, "_check_model", new=mocker.AsyncMock(return_value=None))
+        mocker.patch.object(
+            speech_server,
+            "_generate_audio_bytes",
+            new=mocker.AsyncMock(side_effect=RuntimeError("codec failed")),
+        )
+
+        response = await speech_server.create_speech(OpenAICreateSpeechRequest(input="Hello"))
+
+        assert isinstance(response, ErrorResponse)
+        assert response.error.code == HTTPStatus.INTERNAL_SERVER_ERROR
+        assert response.error.type == "InternalServerError"
+
     def test_is_tts_detection_no_stage(self, speech_server):
         """Test TTS model detection when no TTS stage exists."""
         # Fixture creates server with stage_configs = [] -> _is_tts should be False
@@ -3925,11 +3964,19 @@ class TestAsyncOmniSupportedTasks:
         assert "generate" in tasks
 
 
-def test_api_server_create_speech_wraps_error_response_status(mocker: MockerFixture):
+@pytest.mark.parametrize(
+    ("status_code", "err_type"),
+    [(HTTPStatus.BAD_REQUEST, "BadRequestError"), (HTTPStatus.INTERNAL_SERVER_ERROR, "InternalServerError")],
+)
+def test_api_server_create_speech_wraps_error_response_status(
+    mocker: MockerFixture,
+    status_code: HTTPStatus,
+    err_type: str,
+):
     handler = mocker.MagicMock()
     handler.create_speech = mocker.AsyncMock(
         return_value=ErrorResponse(
-            error=ErrorInfo(message="bad request", type="BadRequestError", param=None, code=400),
+            error=ErrorInfo(message="speech failed", type=err_type, param=None, code=status_code),
         )
     )
 
@@ -3938,7 +3985,12 @@ def test_api_server_create_speech_wraps_error_response_status(mocker: MockerFixt
 
     response = asyncio.run(api_server_module.create_speech(request, raw_request))
 
-    _assert_openai_error_response(response, status_code=400, message="bad request")
+    _assert_openai_error_response(
+        response,
+        status_code=status_code,
+        message="speech failed",
+        err_type=err_type,
+    )
 
 
 def _make_api_server_request(handler, *, method: str = "POST", path: str = "/v1/audio/voices") -> Request:
diff --git a/vllm_omni/entrypoints/openai/serving_speech.py b/vllm_omni/entrypoints/openai/serving_speech.py
index c044a67b6..40ae2f911 100644
--- a/vllm_omni/entrypoints/openai/serving_speech.py
+++ b/vllm_omni/entrypoints/openai/serving_speech.py
@@ -2726,8 +2726,11 @@ class OmniOpenAIServingSpeech(OpenAIServing, AudioMixin):
                 total_ms,
                 e,
             )
-            logger.exception("Speech generation failed: %s", e)
-            return self.create_error_response(f"Speech generation failed: {e}")
+            return self.create_error_response(
+                f"Speech generation failed: {e}",
+                err_type="InternalServerError",
+                status_code=HTTPStatus.INTERNAL_SERVER_ERROR,
+            )
 
     @staticmethod
     def _merge_batch_item(