feat: GenAI - Added audio_timestamp to GenerationConfig.

PiperOrigin-RevId: 689902378

tests/system/vertexai/test_generative_models.py

@@ -344,6 +344,18 @@ def test_generate_content_from_text_and_remote_video(self, api_endpoint_env_name
         assert response.text
         assert "Zootopia" in response.text
 
+    def test_generate_content_from_text_and_remote_audio(self, api_endpoint_env_name):
+        vision_model = generative_models.GenerativeModel(GEMINI_VISION_MODEL_NAME)
+        audio_part = generative_models.Part.from_uri(
+            uri="gs://cloud-samples-data/audio/speech_16k.wav",
+            mime_type="audio/wav",
+        )
+        response = vision_model.generate_content(
+            contents=["What is in the audio?", audio_part],
+            generation_config=generative_models.GenerationConfig(audio_timestamp=True),
+        )
+        assert response.text
+
     def test_grounding_google_search_retriever(self, api_endpoint_env_name):
         model = preview_generative_models.GenerativeModel(GEMINI_MODEL_NAME)
         google_search_retriever_tool = (

tests/unit/vertexai/test_generative_models.py

@@ -761,6 +761,33 @@ def test_generate_content(self, generative_models: generative_models):
         )
         assert response4.text
 
+        model5 = generative_models.GenerativeModel("gemini-1.5-pro-002")
+        response5 = model5.generate_content(
+            contents=[
+                generative_models.Part.from_uri(
+                    "gs://cloud-samples-data/generative-ai/audio/pixel.mp3",
+                    mime_type="audio/mpeg",
+                ),
+                "What is the audio about?",
+            ],
+            generation_config=generative_models.GenerationConfig(
+                audio_timestamp=True,
+            ),
+            safety_settings=[
+                generative_models.SafetySetting(
+                    category=generative_models.SafetySetting.HarmCategory.HARM_CATEGORY_DANGEROUS_CONTENT,
+                    threshold=generative_models.SafetySetting.HarmBlockThreshold.BLOCK_MEDIUM_AND_ABOVE,
+                    method=generative_models.SafetySetting.HarmBlockMethod.SEVERITY,
+                ),
+                generative_models.SafetySetting(
+                    category=generative_models.SafetySetting.HarmCategory.HARM_CATEGORY_HATE_SPEECH,
+                    threshold=generative_models.SafetySetting.HarmBlockThreshold.BLOCK_ONLY_HIGH,
+                    method=generative_models.SafetySetting.HarmBlockMethod.PROBABILITY,
+                ),
+            ],
+        )
+        assert response5.text
+
     @mock.patch.object(
         target=prediction_service.PredictionServiceClient,
         attribute="generate_content",

vertexai/generative_models/_generative_models.py

@@ -1683,6 +1683,7 @@ def __init__(
         response_mime_type: Optional[str] = None,
         response_schema: Optional[Dict[str, Any]] = None,
         seed: Optional[int] = None,
+        audio_timestamp: Optional[bool] = None,
         routing_config: Optional["RoutingConfig"] = None,
         logprobs: Optional[int] = None,
         response_logprobs: Optional[bool] = None,
@@ -1712,6 +1713,7 @@ def __init__(
                 The model needs to be prompted to output the appropriate
                 response type, otherwise the behavior is undefined.
             response_schema: Output response schema of the genreated candidate text.
+            audio_timestamp: If true, the timestamp of the audio will be included in the response.
             routing_config: Model routing preference set in the request.
             logprobs: Logit probabilities.
             reponse_logprobs: If true, export the logprobs results in response.
@@ -1728,6 +1730,7 @@ def __init__(
                     max_output_tokens=100,
                     stop_sequences=["\n\n\n"],
                     seed=5,
+                    audio_timestamp=True,
                 )
             )
             ```
@@ -1750,6 +1753,7 @@ def __init__(
             response_mime_type=response_mime_type,
             response_schema=raw_schema,
             seed=seed,
+            audio_timestamp=audio_timestamp,
             logprobs=logprobs,
             response_logprobs=response_logprobs,
         )