diff --git a/src/agents/voice/model.py b/src/agents/voice/model.py
index 3b4a8e85b5..c4dce67e96 100644
--- a/src/agents/voice/model.py
+++ b/src/agents/voice/model.py
@@ -87,6 +87,12 @@ class TTSModelSettings:
     speed: float | None = None
     """The speed with which the TTS model will read the text. Between 0.25 and 4.0."""
 
+    def __post_init__(self) -> None:
+        # Configurations loaded from JSON/YAML commonly represent NumPy dtypes as strings.
+        # Normalize those spellings once at the settings boundary so downstream consumers can
+        # compare against the supported NumPy dtypes consistently.
+        self.dtype = np.dtype(self.dtype)
+
 
 class TTSModel(abc.ABC):
     """A text-to-speech model that can convert text into audio output."""
@@ -228,3 +234,4 @@ def get_stt_model(self, model_name: str | None) -> STTModel:
     @abc.abstractmethod
     def get_tts_model(self, model_name: str | None) -> TTSModel:
         """Get a text-to-speech model by name."""
+        pass
diff --git a/tests/voice/test_tts_model_settings.py b/tests/voice/test_tts_model_settings.py
new file mode 100644
index 0000000000..db2470c234
--- /dev/null
+++ b/tests/voice/test_tts_model_settings.py
@@ -0,0 +1,23 @@
+from __future__ import annotations
+
+import numpy as np
+
+from agents.voice import TTSModelSettings
+
+
+def test_tts_model_settings_normalizes_string_dtype() -> None:
+    settings = TTSModelSettings(dtype="float32")
+
+    assert settings.dtype == np.dtype("float32")
+
+
+def test_tts_model_settings_normalizes_int16_dtype() -> None:
+    settings = TTSModelSettings(dtype="int16")
+
+    assert settings.dtype == np.dtype("int16")
+
+
+def test_tts_model_settings_accepts_numpy_dtype() -> None:
+    settings = TTSModelSettings(dtype=np.float32)
+
+    assert settings.dtype == np.dtype("float32")