From 428c9f6bf38730518d8ff13273a8d88445f2becc Mon Sep 17 00:00:00 2001
From: Fahima Mokhtari <df_mokhtari@esi.dz>
Date: Fri, 23 Oct 2026 21:10:51 -0400
Subject: [PATCH] fix(aws_polly): read the engine from the bare model name
 (#47963)

* fix(aws_polly): read the engine from the bare model name

speech() strips the aws_polly/ prefix before AWSPollyTextToSpeechConfig sees
the model, so _extract_engine_from_model got a bare "standard", found no
slash and returned the neural default. Every Polly request went out with
Engine: neural whatever the model named.

Take the last path segment instead, which covers both the bare name and the
prefixed one.

* test(aws_polly): move imports to the top of the file

* docs(aws_polly): update the engine comment to the bare model name
---
 .../text_to_speech/transformation.py          | 17 +++++----
 ...aws_polly_text_to_speech_transformation.py | 36 +++++++++++++++++--
 .../text_to_speech/test_transformation.py     | 16 +++++++++
 3 files changed, 58 insertions(+), 11 deletions(-)

diff --git a/litellm/llms/aws_polly/text_to_speech/transformation.py b/litellm/llms/aws_polly/text_to_speech/transformation.py
index 01fec02c310..325c9fd364d 100644
--- a/litellm/llms/aws_polly/text_to_speech/transformation.py
+++ b/litellm/llms/aws_polly/text_to_speech/transformation.py
@@ -178,7 +178,7 @@ class AWSPollyTextToSpeechConfig(BaseTextToSpeechConfig, BaseAWSLLM):
         else:
             mapped_params["output_format"] = self.DEFAULT_OUTPUT_FORMAT
 
-        # Extract engine from model name (e.g., "aws_polly/neural" -> "neural")
+        # model arrives without the "aws_polly/" prefix, e.g. "standard"
         engine: Final = self._extract_engine_from_model(model)
         mapped_params["engine"] = engine
 
@@ -196,18 +196,17 @@ class AWSPollyTextToSpeechConfig(BaseTextToSpeechConfig, BaseAWSLLM):
         """
         Extract engine from model name.
 
+        speech() strips the "aws_polly/" prefix before the config is called,
+        so the bare engine name is what normally arrives here.
+
         Examples:
-            - aws_polly/neural -> neural
-            - aws_polly/standard -> standard
+            - standard -> standard
             - aws_polly/long-form -> long-form
             - aws_polly -> neural (default)
         """
-        if "/" in model:
-            parts: Final = model.split("/")
-            if len(parts) >= 2:
-                engine: Final = parts[1].lower()
-                if engine in self.VALID_ENGINES:
-                    return engine
+        engine: Final = model.rsplit("/", 1)[-1].lower()
+        if engine in self.VALID_ENGINES:
+            return engine
         return self.DEFAULT_ENGINE
 
     def validate_environment(
diff --git a/tests/unit/llms/aws_polly/text_to_speech/test_aws_polly_text_to_speech_transformation.py b/tests/unit/llms/aws_polly/text_to_speech/test_aws_polly_text_to_speech_transformation.py
index 3826b22af34..ace82e517ad 100644
--- a/tests/unit/llms/aws_polly/text_to_speech/test_aws_polly_text_to_speech_transformation.py
+++ b/tests/unit/llms/aws_polly/text_to_speech/test_aws_polly_text_to_speech_transformation.py
@@ -1,6 +1,10 @@
-import litellm
+import json
+from unittest.mock import MagicMock, patch
+
+import httpx
 import pytest
-from unittest.mock import MagicMock
+
+import litellm
 
 
 @pytest.mark.asyncio
@@ -134,6 +138,34 @@ async def test_aws_polly_tts_with_native_voice():
         assert request_body.get("TextType", "text") == "text"
 
 
+@pytest.mark.asyncio
+@pytest.mark.usefixtures("fake_provider_credentials")
+@pytest.mark.parametrize("engine", ["standard", "neural", "long-form", "generative"])
+async def test_aws_polly_tts_sends_engine_from_model(engine: str) -> None:
+    """
+    The engine in the model name must reach Polly. speech() strips the
+    provider prefix before the config reads it, which used to turn every
+    engine into the neural default.
+    """
+    mock_httpx_response = MagicMock(spec=httpx.Response)
+    mock_httpx_response.content = b"fake_audio_data"
+    mock_httpx_response.status_code = 200
+    mock_httpx_response.headers = {"content-type": "audio/mpeg"}
+
+    with patch("litellm.llms.custom_httpx.http_handler.AsyncHTTPHandler.post") as mock_post:
+        mock_post.return_value = mock_httpx_response
+
+        await litellm.aspeech(
+            model=f"aws_polly/{engine}",
+            voice="Joanna",
+            input="Hello",
+            aws_region_name="us-east-1",
+        )
+
+        request_body = json.loads(mock_post.call_args.kwargs["data"])
+        assert request_body["Engine"] == engine
+
+
 @pytest.mark.asyncio
 @pytest.mark.usefixtures("fake_provider_credentials")
 async def test_aws_polly_tts_with_openai_voice_mapping():
diff --git a/tests/unit/llms/aws_polly/text_to_speech/test_transformation.py b/tests/unit/llms/aws_polly/text_to_speech/test_transformation.py
index 4778b6296fc..5f195b7a478 100644
--- a/tests/unit/llms/aws_polly/text_to_speech/test_transformation.py
+++ b/tests/unit/llms/aws_polly/text_to_speech/test_transformation.py
@@ -1,6 +1,8 @@
 import json
 from typing import Final
 
+import pytest
+
 from litellm.llms.aws_polly.text_to_speech.transformation import AWSPollyTextToSpeechConfig
 
 
@@ -17,3 +19,17 @@ def test_installed_botocore_signs_the_speech_request() -> None:
     authorization: Final = headers["Authorization"]
     assert "/us-west-2/polly/aws4_request" in authorization
     assert json.loads(body) == {"Text": "ping", "VoiceId": "Joanna"}
+
+
+@pytest.mark.parametrize(
+    ("model", "expected"),
+    [
+        ("standard", "standard"),
+        ("Long-Form", "long-form"),
+        ("aws_polly/generative", "generative"),
+        ("aws_polly", "neural"),
+        ("unknown", "neural"),
+    ],
+)
+def test_extract_engine_from_model(model: str, expected: str) -> None:
+    assert AWSPollyTextToSpeechConfig()._extract_engine_from_model(model) == expected