diff --git a/Dockerfile b/Dockerfile
index 4bcefafd..eb93f634 100644
--- a/Dockerfile
+++ b/Dockerfile
@@ -28,4 +28,4 @@ RUN uv sync --extra api --extra cpu --no-dev \
   && uv run glados download
 
 EXPOSE 5050
-CMD ["uv", "run", "litestar", "--app", "glados.api.app:app", "run", "--host", "0.0.0.0", "--port", "5050"]
+CMD ["uv", "run", "litestar", "--app", "glados.api.app:create_app", "run", "--host", "0.0.0.0", "--port", "5050"]
diff --git a/README.md b/README.md
index 4da905fa..3f9e03c2 100644
--- a/README.md
+++ b/README.md
@@ -506,6 +506,19 @@ Or Docker:
 docker compose up -d --build
 ```
 
+By default the API reuses one `SpeechSynthesizer` across requests (`reuse_tts: true` in `configs/api_config.yaml`). Disable it to restore per-request model loading:
+
+```yaml
+Api:
+  reuse_tts: false
+```
+
+Environment override:
+
+```bash
+GLADOS_API_REUSE_TTS=false ./scripts/serve
+```
+
 Generate speech:
 ```bash
 curl -X POST http://localhost:5050/v1/audio/speech \
diff --git a/configs/api_config.yaml b/configs/api_config.yaml
new file mode 100644
index 00000000..8562e5d7
--- /dev/null
+++ b/configs/api_config.yaml
@@ -0,0 +1,3 @@
+Api:
+  # Reuse one SpeechSynthesizer across requests instead of reloading ONNX each call.
+  reuse_tts: true
diff --git a/scripts/serve b/scripts/serve
index d71315cd..98a14ad9 100755
--- a/scripts/serve
+++ b/scripts/serve
@@ -2,4 +2,4 @@
 # Start the API server.
 set -e
 
-uv run litestar --app glados.api.app:app run --host 0.0.0.0 --port 5050 --debug --reload
+uv run litestar --app glados.api.app:create_app run --host 0.0.0.0 --port 5050 --debug --reload
diff --git a/src/glados/api/app.py b/src/glados/api/app.py
index 28a355dd..d248cce0 100644
--- a/src/glados/api/app.py
+++ b/src/glados/api/app.py
@@ -1,16 +1,25 @@
+from collections.abc import AsyncIterator
+from contextlib import asynccontextmanager
 from dataclasses import dataclass
 import io
+from pathlib import Path
 from typing import Literal
 
 from litestar import Litestar, post
 from litestar.response import Stream
+import yaml
 
+from glados.utils.resources import resource_path
+
+from .config import ApiConfig
 from .log import structlog_plugin
-from .tts import write_glados_audio_file
+from .tts import configure_tts, write_glados_audio_file
 
 Voice = Literal["glados"]
 ResponseFormat = Literal["mp3", "wav", "ogg"]
 
+DEFAULT_API_CONFIG = resource_path("configs/api_config.yaml")
+
 
 @dataclass
 class RequestData:
@@ -24,6 +33,16 @@ class RequestData:
 CONTENT_TYPES: dict[ResponseFormat, str] = {"mp3": "audio/mpeg", "wav": "audio/wav", "ogg": "audio/ogg"}
 
 
+def load_api_config(path: str | Path | None = None) -> ApiConfig:
+    """Load API settings from configs/api_config.yaml (Api: section)."""
+    config_path = Path(path) if path is not None else DEFAULT_API_CONFIG
+    data: dict[str, object] = {}
+    if config_path.exists():
+        raw = yaml.safe_load(config_path.read_text(encoding="utf-8")) or {}
+        data = raw.get("Api", {}) or {}
+    return ApiConfig.model_validate(data)
+
+
 @post("/v1/audio/speech")
 async def create_speech(data: RequestData) -> Stream:
     """
@@ -49,4 +68,15 @@ async def create_speech(data: RequestData) -> Stream:
     )
 
 
-app = Litestar([create_speech], plugins=[structlog_plugin])
+def create_app() -> Litestar:
+    """Create the Litestar application for the TTS API server."""
+
+    @asynccontextmanager
+    async def lifespan(app: Litestar) -> AsyncIterator[None]:
+        configure_tts(load_api_config())
+        yield
+
+    return Litestar([create_speech], plugins=[structlog_plugin], lifespan=[lifespan])
+
+
+app = create_app()
diff --git a/src/glados/api/config.py b/src/glados/api/config.py
new file mode 100644
index 00000000..ab73addd
--- /dev/null
+++ b/src/glados/api/config.py
@@ -0,0 +1,19 @@
+import os
+
+from pydantic import BaseModel, Field, model_validator
+
+
+class ApiConfig(BaseModel):
+    """Configuration for the OpenAI-compatible TTS API server."""
+
+    reuse_tts: bool = Field(
+        default=True,
+        description="Reuse a single SpeechSynthesizer instance across requests instead of reloading ONNX each call.",
+    )
+
+    @model_validator(mode="after")
+    def _apply_env_overrides(self) -> "ApiConfig":
+        env_reuse_tts = os.environ.get("GLADOS_API_REUSE_TTS")
+        if env_reuse_tts is not None:
+            self.reuse_tts = env_reuse_tts.lower() in ("1", "true", "yes")
+        return self
diff --git a/src/glados/api/tts.py b/src/glados/api/tts.py
index f03120c1..9a2486e3 100644
--- a/src/glados/api/tts.py
+++ b/src/glados/api/tts.py
@@ -1,10 +1,39 @@
+from functools import lru_cache
 import io
 
 import soundfile as sf
 
-from glados.TTS import tts_glados
+from glados.TTS import SpeechSynthesizerProtocol, get_speech_synthesizer
 from glados.utils import spoken_text_converter
 
+from .config import ApiConfig
+
+_api_config = ApiConfig()
+
+
+def configure_tts(config: ApiConfig) -> None:
+    """Apply API TTS settings and warm models when reuse is enabled."""
+    global _api_config
+    _api_config = config
+    if config.reuse_tts:
+        warm_tts()
+
+
+@lru_cache(maxsize=1)
+def _get_synthesizer() -> SpeechSynthesizerProtocol:
+    return get_speech_synthesizer("glados")
+
+
+@lru_cache(maxsize=1)
+def _get_text_converter() -> spoken_text_converter.SpokenTextConverter:
+    return spoken_text_converter.SpokenTextConverter()
+
+
+def warm_tts() -> None:
+    """Load ONNX session and text converter at startup."""
+    _get_synthesizer()
+    _get_text_converter()
+
 
 def write_glados_audio_file(f: str | io.BytesIO, text: str, *, format: str) -> None:
     """Generate GLaDOS-style speech audio from text and write to a file.
@@ -14,9 +43,13 @@ def write_glados_audio_file(f: str | io.BytesIO, text: str, *, format: str) -> N
         text: Text to convert to speech
         format: Audio format (e.g., "mp3", "wav", "ogg")
     """
-    glados_tts = tts_glados.SpeechSynthesizer()
-    converter = spoken_text_converter.SpokenTextConverter()
-    converted_text = converter.text_to_spoken(text)
+    if _api_config.reuse_tts:
+        glados_tts = _get_synthesizer()
+        converted_text = _get_text_converter().text_to_spoken(text)
+    else:
+        glados_tts = get_speech_synthesizer("glados")
+        converted_text = spoken_text_converter.SpokenTextConverter().text_to_spoken(text)
+
     audio = glados_tts.generate_speech_audio(converted_text)
     sf.write(
         f,
diff --git a/tests/test_api_config.py b/tests/test_api_config.py
new file mode 100644
index 00000000..31c42e38
--- /dev/null
+++ b/tests/test_api_config.py
@@ -0,0 +1,34 @@
+from pathlib import Path
+
+import pytest
+
+from glados.api.app import load_api_config
+from glados.api.config import ApiConfig
+
+
+def test_api_config_defaults_to_reuse_tts() -> None:
+    config = ApiConfig()
+    assert config.reuse_tts is True
+
+
+def test_api_config_yaml_section(tmp_path: Path) -> None:
+    config_file = tmp_path / "api_config.yaml"
+    config_file.write_text("Api:\n  reuse_tts: false\n", encoding="utf-8")
+
+    config = load_api_config(str(config_file))
+
+    assert config.reuse_tts is False
+
+
+def test_api_config_env_override(monkeypatch: pytest.MonkeyPatch) -> None:
+    monkeypatch.setenv("GLADOS_API_REUSE_TTS", "false")
+
+    config = ApiConfig.model_validate({})
+
+    assert config.reuse_tts is False
+
+
+def test_load_api_config_uses_defaults_when_file_missing(tmp_path: Path) -> None:
+    config = load_api_config(str(tmp_path / "missing.yaml"))
+
+    assert config.reuse_tts is True