[Test] Add voice or language test case for Qwen3-omni and Qwen-tts (#1844)

Signed-off-by: yenuo26 <410167048@qq.com>
Signed-off-by: wangyu <410167048@qq.com>
Signed-off-by: wangyu <53896905+yenuo26@users.noreply.github.com>

.buildkite/test-merge.yml

@@ -56,8 +56,8 @@ steps:
     timeout_in_minutes: 20
     depends_on: upload-merge-pipeline
     commands:
-      - pytest -s -v tests/e2e/offline_inference/test_diffusion_cpu_offload.py
-      - pytest -s -v tests/e2e/offline_inference/test_diffusion_layerwise_offload.py
+      # Single pytest session for one combined summary at end of log.
+      - pytest -s -v tests/e2e/offline_inference/test_diffusion_cpu_offload.py tests/e2e/offline_inference/test_diffusion_layerwise_offload.py
     agents:
       queue: "gpu_1_queue" # g6.4xlarge instance on AWS, has 1 L4 GPU
     plugins:
@@ -111,8 +111,7 @@ steps:
     timeout_in_minutes: 20
     depends_on: upload-merge-pipeline
     commands:
-      - pytest -s -v tests/e2e/offline_inference/test_sequence_parallel.py
-      - pytest -s -v tests/diffusion/distributed/test_ulysses_uaa_perf.py
+      - pytest -s -v tests/e2e/offline_inference/test_sequence_parallel.py tests/diffusion/distributed/test_ulysses_uaa_perf.py
     agents:
       queue: "gpu_4_queue" # g6.12xlarge instance on AWS, has 4 L4 GPU
     plugins:
@@ -193,8 +192,7 @@ steps:
     commands:
       - export VLLM_LOGGING_LEVEL=DEBUG
       - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-      - pytest -s -v tests/e2e/offline_inference/test_qwen2_5_omni.py
-      - pytest -s -v tests/e2e/online_serving/test_qwen2_5_omni.py -m "advanced_model" --run-level "advanced_model"
+      - pytest -s -v tests/e2e/offline_inference/test_qwen2_5_omni.py tests/e2e/online_serving/test_qwen2_5_omni.py -m "advanced_model" --run-level "advanced_model"
     agents:
       queue: "gpu_4_queue" # g6.12xlarge instance on AWS, has 4 L4 GPU
     plugins:
@@ -216,7 +214,7 @@ steps:
           export VLLM_LOGGING_LEVEL=DEBUG
           export VLLM_WORKER_MULTIPROC_METHOD=spawn
           export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
-          pytest -s -v tests/e2e/online_serving/test_qwen3_tts_customvoice.py -m "advanced_model" --run-level "advanced_model" && pytest -s -v tests/e2e/offline_inference/test_qwen3_tts_customvoice.py
+          pytest -s -v tests/e2e/online_serving/test_qwen3_tts_customvoice.py tests/e2e/offline_inference/test_qwen3_tts_customvoice.py -m "advanced_model" --run-level "advanced_model"
         '
     agents:
       queue: "gpu_1_queue"
@@ -239,7 +237,7 @@ steps:
           export VLLM_LOGGING_LEVEL=DEBUG
           export VLLM_WORKER_MULTIPROC_METHOD=spawn
           export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
-          pytest -s -v tests/e2e/online_serving/test_qwen3_tts_base.py -m "advanced_model" --run-level "advanced_model" && pytest -s -v tests/e2e/offline_inference/test_qwen3_tts_base.py
+          pytest -s -v tests/e2e/online_serving/test_qwen3_tts_base.py tests/e2e/offline_inference/test_qwen3_tts_base.py -m "advanced_model" --run-level "advanced_model"
         '
     agents:
       queue: "gpu_1_queue"
@@ -259,9 +257,8 @@ steps:
     depends_on: upload-merge-pipeline
     commands:
       - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-      - pytest -s -v tests/e2e/offline_inference/test_qwen3_omni.py
-      - pytest -s -v tests/e2e/online_serving/test_qwen3_omni.py -m "advanced_model" --run-level "advanced_model"
-      - pytest -s -v tests/e2e/online_serving/test_mimo_audio.py -m "advanced_model" --run-level "advanced_model"
+      - export VLLM_TEST_CLEAN_GPU_MEMORY="1"
+      - pytest -s -v tests/e2e/offline_inference/test_qwen3_omni.py tests/e2e/online_serving/test_qwen3_omni.py tests/e2e/online_serving/test_mimo_audio.py -m "advanced_model" --run-level "advanced_model"
     agents:
       queue: "mithril-h100-pool"
     plugins:
@@ -347,8 +344,7 @@ steps:
           export VLLM_TEST_CLEAN_GPU_MEMORY=1
           export VLLM_IMAGE_FETCH_TIMEOUT=60
           pytest -s -v tests/e2e/offline_inference/test_bagel_text2img.py -m "advanced_model" --run-level "advanced_model" -k "shared_memory"
-          pytest -s -v tests/e2e/offline_inference/test_bagel_img2img.py -m "advanced_model" --run-level "advanced_model"
-          pytest -s -v tests/e2e/online_serving/test_bagel_online.py -m "advanced_model" --run-level "advanced_model"
+          pytest -s -v tests/e2e/offline_inference/test_bagel_img2img.py tests/e2e/online_serving/test_bagel_online.py -m "advanced_model" --run-level "advanced_model"
         '
     agents:
       queue: "mithril-h100-pool"
@@ -392,8 +388,7 @@ steps:
         timeout 20m bash -c '
           export VLLM_LOGGING_LEVEL=DEBUG
           export VLLM_WORKER_MULTIPROC_METHOD=spawn
-          pytest -s -v tests/e2e/online_serving/test_voxtral_tts.py -m "advanced_model" --run-level "advanced_model"
-          pytest -s -v tests/e2e/offline_inference/test_voxtral_tts.py -m "advanced_model" --run-level "advanced_model"
+          pytest -s -v tests/e2e/online_serving/test_voxtral_tts.py tests/e2e/offline_inference/test_voxtral_tts.py -m "advanced_model" --run-level "advanced_model"
         '
     agents:
       queue: "mithril-h100-pool"

.buildkite/test-nightly.yml

@@ -6,16 +6,10 @@ steps:
     commands:
       - export VLLM_WORKER_MULTIPROC_METHOD=spawn
       - |
-        set +e
-        pytest -s -v tests/e2e/online_serving/test_qwen3_omni_expansion.py -m "advanced_model" --run-level "advanced_model"
-        EXIT1=$$?
-        pytest -s -v tests/examples/online_serving/test_qwen3_omni.py -m "advanced_model" --run-level "advanced_model"
-        EXIT2=$$?
-        pytest -s -v tests/e2e/online_serving/test_voxtral_tts.py -m "advanced_model" --run-level "advanced_model"
-        EXIT3=$$?
-        pytest -s -v tests/e2e/offline_inference/test_voxtral_tts.py -m "advanced_model" --run-level "advanced_model"
-        EXIT4=$$?
-        exit $$((EXIT1 | EXIT2 | EXIT3 | EXIT4))
+        pytest -s -v \
+          tests/examples/ \
+          tests/e2e/online_serving/test_*_expansion.py \
+          -m "advanced_model and H100 and omni" --run-level "advanced_model"
     agents:
       queue: "mithril-h100-pool"
     plugins:
@@ -57,8 +51,7 @@ steps:
     commands:
       - export VLLM_WORKER_MULTIPROC_METHOD=spawn
       - export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
-      - pytest -s -v tests/examples/ -m "advanced_model and L4 and omni" --run-level "advanced_model"
-      - pytest -s -v tests/e2e/online_serving/test_*_expansion.py -m "advanced_model and L4 and omni" --run-level "advanced_model"
+      - pytest -s -v tests/examples/ tests/e2e/online_serving/test_*_expansion.py -m "advanced_model and L4 and omni" --run-level "advanced_model"
     agents:
       queue: "gpu_4_queue" # g6.12xlarge instance on AWS, has 4 L4 GPU
     plugins:

.buildkite/test-ready.yml

@@ -328,8 +328,7 @@ steps:
         timeout 20m bash -c '
           export VLLM_LOGGING_LEVEL=DEBUG
           export VLLM_WORKER_MULTIPROC_METHOD=spawn
-          pytest -s -v tests/e2e/online_serving/test_voxtral_tts.py -m "advanced_model" --run-level "advanced_model"
-          pytest -s -v tests/e2e/offline_inference/test_voxtral_tts.py -m "advanced_model" --run-level "advanced_model"
+          pytest -s -v tests/e2e/online_serving/test_voxtral_tts.py -m "core_model" --run-level "core_model"
         '
     agents:
       queue: "mithril-h100-pool"

tests/conftest.py

@@ -1146,18 +1146,6 @@ def convert_audio_bytes_to_text(raw_bytes: bytes) -> str:
     return text
 
 
-def merge_base64_and_convert_to_text(base64_list):
-    """
-    Merge a list of base64 encoded audio data and convert to text.
-    """
-    merged_audio = _merge_base64_audio_to_segment(base64_list)
-    output_path = f"./test_{uuid.uuid4().hex}.wav"
-    merged_audio.export(output_path, format="wav")
-    print(f"audio data is saved: {output_path}")
-    text = convert_audio_file_to_text(output_path)
-    return text
-
-
 def modify_stage_config(
     yaml_path: str,
     updates: dict[str, Any] = None,
@@ -1742,7 +1730,7 @@ def _estimate_voice_gender_from_audio(audio_bytes: bytes) -> str:
         label = str(top.get("label", "")).lower()
         conf = float(top.get("score", 0.0))
 
-        if conf < 0.6:
+        if conf < 0.5:
             gender = "unknown"
         # Some models use non-English labels (e.g., Russian). Normalize to 'male'/'female'.
         elif ("female" in label) or ("жен" in label):
@@ -1771,6 +1759,34 @@ def _estimate_voice_gender_from_audio(audio_bytes: bytes) -> str:
         return "unknown"
 
 
+_PRESET_VOICE_GENDER_MAP: dict[str, str] = {
+    "serena": "female",
+    "uncle_fu": "male",
+    "chelsie": "female",
+    "clone": "female",
+    "ethan": "male",
+}
+
+
+def _assert_preset_voice_gender_from_audio(
+    audio_bytes: bytes | None,
+    voice_name: str | None,
+) -> None:
+    """If ``voice_name`` matches a known preset, assert classifier gender matches (skip when unknown)."""
+    if not voice_name or not audio_bytes:
+        return
+    key = str(voice_name).lower()
+    expected_gender = _PRESET_VOICE_GENDER_MAP.get(key)
+    if expected_gender is None:
+        return
+    estimated_gender = _estimate_voice_gender_from_audio(audio_bytes)
+    print(f"Preset voice gender check: preset={key!r}, estimated={estimated_gender!r}, expected={expected_gender!r}")
+    if estimated_gender != "unknown":
+        assert estimated_gender == expected_gender, (
+            f"{voice_name!r} is expected {expected_gender}, but estimated gender is {estimated_gender!r}"
+        )
+
+
 # Threshold aligned with _compute_pcm_hnr_db docstring (clean clone vs distorted).
 _MIN_PCM_SPEECH_HNR_DB = 1.0
 
@@ -1837,6 +1853,12 @@ def assert_omni_response(response: OmniResponse, request_config: dict[str, Any],
         if "audio" in modalities:
             assert response.audio_content is not None, "No audio output is generated"
             print(f"audio content is: {response.audio_content}")
+            speaker = request_config.get("speaker")
+            if speaker:
+                _assert_preset_voice_gender_from_audio(
+                    response.audio_bytes,
+                    speaker,
+                )
 
         if "text" in modalities:
             assert response.text_content is not None, "No text output is generated"
@@ -1849,12 +1871,14 @@ def assert_omni_response(response: OmniResponse, request_config: dict[str, Any],
             keywords = keywords_dict.get(word_type)
             if "text" in modalities:
                 if keywords:
-                    assert any(keyword in response.text_content.lower() for keyword in keywords), (
+                    text_lower = response.text_content.lower()
+                    assert any(str(kw).lower() in text_lower for kw in keywords), (
                         "The output does not contain any of the keywords."
                     )
             else:
                 if keywords:
-                    assert any(keyword in response.audio_content.lower() for keyword in keywords), (
+                    audio_lower = response.audio_content.lower()
+                    assert any(str(kw).lower() in audio_lower for kw in keywords), (
                         "The output does not contain any of the keywords."
                     )
 
@@ -1908,24 +1932,12 @@ def assert_audio_speech_response(
                 f"Transcript doesn't match input: similarity={similarity:.2f}, transcript='{transcript}'"
             )
 
-        # Voice gender consistency check:
+        # Voice gender consistency check (preset names in ``_PRESET_VOICE_GENDER_MAP``).
         # When the estimator returns 'unknown', we treat it as inconclusive and do NOT fail the test.
-        voice = (request_config.get("voice") or "").lower()
-        if voice and response.audio_bytes:
-            estimated_gender = _estimate_voice_gender_from_audio(response.audio_bytes)
-            voice_gender_map = {
-                # adjust this mapping to your actual voice names
-                "serena": "female",
-                "uncle_fu": "male",
-                "clone": "female",
-            }
-            expected_gender = voice_gender_map.get(voice)
-            if expected_gender is not None:
-                print(f"Estimated voice gender from audio: {estimated_gender} (voice='{voice}')")
-                if estimated_gender != "unknown":
-                    assert estimated_gender == expected_gender, (
-                        f"Voice '{voice}' is expected {expected_gender}, but estimated gender is '{estimated_gender}'"
-                    )
+        _assert_preset_voice_gender_from_audio(
+            response.audio_bytes,
+            request_config.get("voice"),
+        )
 
 
 def assert_diffusion_response(response: DiffusionResponse, request_config: dict[str, Any], run_level: str = None):
@@ -2041,7 +2053,11 @@ class OpenAIClientHandler:
 
             if audio_data or text_content:
                 if audio_data:
-                    audio_content = merge_base64_and_convert_to_text(audio_data)
+                    merged_seg = _merge_base64_audio_to_segment(audio_data)
+                    wav_buf = BytesIO()
+                    merged_seg.export(wav_buf, format="wav")
+                    result.audio_bytes = wav_buf.getvalue()
+                    audio_content = convert_audio_bytes_to_text(result.audio_bytes)
                 if audio_content and text_content:
                     similarity = cosine_similarity_text(audio_content.lower(), text_content.lower())
 
@@ -2096,7 +2112,8 @@ class OpenAIClientHandler:
 
             if audio_data or text_content:
                 if audio_data:
-                    audio_content = convert_audio_to_text(audio_data)
+                    result.audio_bytes = base64.b64decode(audio_data)
+                    audio_content = convert_audio_bytes_to_text(result.audio_bytes)
                 if audio_content and text_content:
                     similarity = cosine_similarity_text(audio_content.lower(), text_content.lower())
 
@@ -2265,8 +2282,9 @@ class OpenAIClientHandler:
             request_config: Request configuration dictionary containing parameters like model, messages, stream.
                 Optional ``use_audio_in_video`` (bool): when true, sets
                 ``extra_body["mm_processor_kwargs"] = {"use_audio_in_video": True}`` for Qwen-Omni video+audio
-                extraction (merged with any existing ``extra_body`` / ``mm_processor_kwargs``).
-                Optional ``extra_body`` (dict): passed through to ``chat.completions.create`` after merge.
+                extraction.
+                Optional top-level ``speaker`` (str): Qwen3-Omni preset TTS speaker name; sent as
+                ``extra_body["speaker"]`` to ``chat.completions.create``.
             request_num: Number of requests, defaults to 1 (single request)
 
         Returns:
@@ -2278,9 +2296,8 @@ class OpenAIClientHandler:
         modalities = request_config.get("modalities", ["text", "audio"])
 
         extra_body: dict[str, Any] = {}
-        raw_extra = request_config.get("extra_body")
-        if raw_extra:
-            extra_body.update(raw_extra)
+        if "speaker" in request_config:
+            extra_body["speaker"] = request_config["speaker"]
         if request_config.get("use_audio_in_video"):
             mm = dict(extra_body.get("mm_processor_kwargs") or {})
             mm["use_audio_in_video"] = True
@@ -2312,12 +2329,15 @@ class OpenAIClientHandler:
             # Send concurrent requests: run create + process in worker so e2e_latency includes full round-trip.
             def _one_omni_request():
                 start = time.perf_counter()
-                chat_completion = self.client.chat.completions.create(
-                    model=request_config.get("model"),
-                    messages=request_config.get("messages"),
-                    modalities=modalities,
-                    stream=stream,
-                )
+                worker_kwargs: dict[str, Any] = {
+                    "model": request_config.get("model"),
+                    "messages": request_config.get("messages"),
+                    "modalities": modalities,
+                    "stream": stream,
+                }
+                if extra_body_arg is not None:
+                    worker_kwargs["extra_body"] = extra_body_arg
+                chat_completion = self.client.chat.completions.create(**worker_kwargs)
                 if stream:
                     response = self._process_stream_omni_response(chat_completion)
                 else:

tests/e2e/offline_inference/test_qwen2_5_omni.py

@@ -57,7 +57,7 @@ def get_question(prompt_type="mix"):
     return prompts.get(prompt_type, prompts["mix"])
 
 
-@pytest.mark.core_model
+@pytest.mark.advanced_model
 @pytest.mark.omni
 @hardware_test(res={"cuda": "L4", "rocm": "MI325", "xpu": "B60"}, num_cards={"cuda": 4, "rocm": 2, "xpu": 3})
 @pytest.mark.parametrize("omni_runner", test_params, indirect=True)
@@ -88,7 +88,7 @@ def test_mix_to_audio(omni_runner, omni_runner_handler) -> None:
     omni_runner_handler.send_request(request_config)
 
 
-@pytest.mark.core_model
+@pytest.mark.advanced_model
 @pytest.mark.omni
 @hardware_test(res={"cuda": "L4", "rocm": "MI325", "xpu": "B60"}, num_cards={"cuda": 4, "rocm": 2, "xpu": 3})
 @pytest.mark.parametrize("omni_runner", test_params, indirect=True)

tests/e2e/offline_inference/test_qwen3_omni.py

@@ -56,7 +56,7 @@ def get_question(prompt_type="video"):
     return prompts.get(prompt_type, prompts["video"])
 
 
-@pytest.mark.core_model
+@pytest.mark.advanced_model
 @pytest.mark.omni
 @hardware_test(res={"cuda": "H100", "rocm": "MI325"}, num_cards=2)
 @pytest.mark.parametrize("omni_runner", test_params, indirect=True)

tests/e2e/online_serving/test_qwen3_omni_expansion.py

@@ -26,7 +26,7 @@ from tests.utils import hardware_test
 model = "Qwen/Qwen3-Omni-30B-A3B-Instruct"
 
 AUDIO_KEY = ["test"]
-IMAGE_KEY = ["square", "quadrate"]
+IMAGE_KEY = ["square", "quadrate", "rectangle"]
 VIDEO_KEY = ["sphere", "globe", "circle", "round", "ball"]
 
 
@@ -103,6 +103,7 @@ def get_prompt(prompt_type="text_only"):
         "text_audio": "What is in this audio? ",
         "text_audio_video": "First, what is in this audio? Then, what is in this video? ",
         "one_word": "What is the capital of UK? Answer in one word",
+        "text_chinese": "北京，中国的首都，是一座融合了长城等历史地点与现代建筑的国际化大都市，充满了独特的文化与活力。请重复这句话。",
     }
     return prompts.get(prompt_type, prompts["text_only"])
 
@@ -464,20 +465,10 @@ def test_audio_in_video_002(omni_server, openai_client) -> None:
         "messages": messages,
         "stream": True,
         "use_audio_in_video": True,
-        "key_words": {"video": VIDEO_KEY, "audio": AUDIO_KEY + ["beep", "electronic"]},
+        "key_words": {"video": VIDEO_KEY},
     }
 
-    # Retry when assert_omni_response fails on key_words (see tests/conftest.py).
-    _keyword_assert_msg = "The output does not contain any of the keywords."
-    _max_retries = 3
-    for attempt in range(_max_retries):
-        try:
-            openai_client.send_omni_request(request_config, request_num=get_max_batch_size())
-            break
-        except AssertionError as e:
-            if _keyword_assert_msg not in str(e) or attempt == _max_retries - 1:
-                raise
-            print(f"Keyword assertion failed, retrying {attempt + 2}/{_max_retries}: {e!r}")
+    openai_client.send_omni_request(request_config, request_num=get_max_batch_size())
 
 
 @pytest.mark.advanced_model
@@ -514,3 +505,120 @@ def test_one_word_prompt_001(omni_server, openai_client) -> None:
             if _similarity_assert_msg not in str(e) or attempt == _max_retries - 1:
                 raise
             print(f"Similarity assertion failed, retrying {attempt + 2}/{_max_retries}: {e!r}")
+
+
+@pytest.mark.advanced_model
+@pytest.mark.omni
+@hardware_test(res={"cuda": "H100", "rocm": "MI325"}, num_cards=2)
+@pytest.mark.parametrize("omni_server", test_params, indirect=True)
+def test_speaker_001(omni_server, openai_client) -> None:
+    """
+    Input Modal: text only (one-word answer constraint).
+    Output Modal: text, audio (default ``modalities``); ``key_words`` only assert on text.
+    Input Setting: stream=True
+    Datasets: single request
+    """
+    messages = dummy_messages_from_mix_data(
+        system_prompt=get_system_prompt(),
+        content_text=get_prompt("text"),
+    )
+
+    request_config = {
+        "model": omni_server.model,
+        "messages": messages,
+        "stream": True,
+        "speaker": "Chelsie",
+        "key_words": {"text": ["beijing"]},
+    }
+
+    openai_client.send_omni_request(request_config)
+
+
+@pytest.mark.advanced_model
+@pytest.mark.omni
+@hardware_test(res={"cuda": "H100", "rocm": "MI325"}, num_cards=2)
+@pytest.mark.parametrize("omni_server", test_params, indirect=True)
+def test_speaker_002(omni_server, openai_client) -> None:
+    """
+    Input Modal: text only (one-word answer constraint).
+    Output Modal: text, audio (default ``modalities``); ``key_words`` only assert on text.
+    Input Setting: stream=True
+    Datasets: single request
+    """
+    messages = dummy_messages_from_mix_data(
+        system_prompt=get_system_prompt(),
+        content_text=get_prompt("text"),
+    )
+
+    request_config = {
+        "model": omni_server.model,
+        "messages": messages,
+        "stream": True,
+        "speaker": "Ethan",
+        "key_words": {"text": ["beijing"]},
+    }
+
+    # Retry only when assert_omni_response fails on preset voice gender (see tests/conftest.py).
+    _gender_assert_substr = "estimated gender"
+    _max_retries = 3
+    for attempt in range(_max_retries):
+        try:
+            openai_client.send_omni_request(request_config, request_num=get_max_batch_size())
+            break
+        except AssertionError as e:
+            if _gender_assert_substr not in str(e) or attempt == _max_retries - 1:
+                raise
+            print(f"Gender assertion failed, retrying {attempt + 2}/{_max_retries}: {e!r}")
+
+
+@pytest.mark.advanced_model
+@pytest.mark.omni
+@hardware_test(res={"cuda": "H100", "rocm": "MI325"}, num_cards=2)
+@pytest.mark.parametrize("omni_server", test_params, indirect=True)
+def test_speaker_003(omni_server, openai_client) -> None:
+    """
+    Input Modal: text only (one-word answer constraint).
+    Output Modal: text, audio (default ``modalities``); ``key_words`` only assert on text.
+    Input Setting: stream=True
+    Datasets: single request
+    """
+    messages = dummy_messages_from_mix_data(
+        system_prompt=get_system_prompt(),
+        content_text=get_prompt("text"),
+    )
+
+    request_config = {
+        "model": omni_server.model,
+        "messages": messages,
+        "stream": True,
+        "speaker": "CHELSIE",
+        "key_words": {"text": ["beijing"]},
+    }
+
+    openai_client.send_omni_request(request_config)
+
+
+@pytest.mark.advanced_model
+@pytest.mark.omni
+@hardware_test(res={"cuda": "H100", "rocm": "MI325"}, num_cards=2)
+@pytest.mark.parametrize("omni_server", test_params, indirect=True)
+def test_language_001(omni_server, openai_client) -> None:
+    """
+    Input Modal: text only (one-word answer constraint).
+    Output Modal: text, audio (default ``modalities``); ``key_words`` only assert on text.
+    Input Setting: stream=True
+    Datasets: single request
+    """
+    messages = dummy_messages_from_mix_data(
+        system_prompt=get_system_prompt(),
+        content_text=get_prompt("text_chinese"),
+    )
+
+    request_config = {
+        "model": omni_server.model,
+        "messages": messages,
+        "stream": True,
+        "key_words": {"text": ["北京"]},
+    }
+
+    openai_client.send_omni_request(request_config)

tests/e2e/online_serving/test_qwen3_tts_customvoice_expansion.py

@@ -120,6 +120,31 @@ def test_voice_002(omni_server, openai_client) -> None:
     openai_client.send_audio_speech_request(request_config)
 
 
+@pytest.mark.advanced_model
+@pytest.mark.omni
+@hardware_test(res={"cuda": "L4"}, num_cards=1)
+@pytest.mark.parametrize("omni_server", tts_server_params, indirect=True)
+def test_voice_003(omni_server, openai_client) -> None:
+    """
+    Test text input processing and audio output via OpenAI API.
+    Deploy Setting: default yaml
+    Input Modal: text
+    Output Modal: audio
+    Input Setting: stream=False, language=chinese
+    Datasets: few requests
+    """
+    request_config = {
+        "model": omni_server.model,
+        "input": get_prompt(),
+        "stream": False,
+        "response_format": "wav",
+        "task_type": "CustomVoice",
+        "voice": "SERENA",
+    }
+
+    openai_client.send_audio_speech_request(request_config)
+
+
 @pytest.mark.advanced_model
 @pytest.mark.omni
 @hardware_test(res={"cuda": "L4"}, num_cards=1)