diff --git a/src/agents/voice/model.py b/src/agents/voice/model.py
index 3b4a8e85b5..079704e56e 100644
--- a/src/agents/voice/model.py
+++ b/src/agents/voice/model.py
@@ -7,6 +7,7 @@
 
 from typing_extensions import TypedDict
 
+from .exceptions import UserError
 from .imports import np, npt
 from .input import AudioInput, StreamedAudioInput
 from .utils import get_sentence_based_splitter
@@ -87,6 +88,15 @@ 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.
+        try:
+            self.dtype = np.dtype(self.dtype)
+        except (TypeError, ValueError) as error:
+            raise UserError("Invalid output dtype") from error
+
 
 class TTSModel(abc.ABC):
     """A text-to-speech model that can convert text into audio output."""
diff --git a/tests/voice/test_tts_model_settings.py b/tests/voice/test_tts_model_settings.py
new file mode 100644
index 0000000000..2b3ea77294
--- /dev/null
+++ b/tests/voice/test_tts_model_settings.py
@@ -0,0 +1,47 @@
+from __future__ import annotations
+
+import numpy as np
+import pytest
+
+from agents.exceptions import UserError
+from agents.voice import AudioInput, TTSModelSettings, VoicePipeline
+
+from .helpers import extract_events
+from .pipeline_test_models import QueuedSTTModel, QueuedVoiceWorkflow, ZeroPcmTTSModel
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    ("dtype", "expected_dtype"),
+    [("int16", np.int16), ("float32", np.float32), ("f4", np.float32)],
+    ids=["int16-string", "float32-string", "float32-alias"],
+)
+async def test_voicepipeline_accepts_string_tts_dtype_from_dictionary_config(
+    dtype: str,
+    expected_dtype: type[np.int16] | type[np.float32],
+) -> None:
+    fake_stt = QueuedSTTModel(["first"])
+    fake_tts = ZeroPcmTTSModel()
+    pipeline = VoicePipeline(
+        workflow=QueuedVoiceWorkflow([["out_1"]]),
+        stt_model=fake_stt,
+        tts_model=fake_tts,
+        config={"tts_settings": {"buffer_size": 1, "dtype": dtype}},
+    )
+
+    result = await pipeline.run(AudioInput(buffer=np.zeros(2, dtype=np.int16)))
+    events, audio_chunks = await extract_events(result)
+
+    assert events == ["turn_started", "audio", "turn_ended", "session_ended"]
+    decoded_audio = np.frombuffer(audio_chunks[0], dtype=expected_dtype)
+    assert decoded_audio.dtype == np.dtype(expected_dtype)
+
+
+@pytest.mark.parametrize(
+    "dtype",
+    ["not-a-dtype", {"names": ["x"], "formats": []}],
+    ids=["unparseable-string", "malformed-structured-dtype"],
+)
+def test_tts_model_settings_preserves_user_error_for_invalid_dtype(dtype: object) -> None:
+    with pytest.raises(UserError, match="Invalid output dtype"):
+        TTSModelSettings(dtype=dtype)  # type: ignore[arg-type]