Make speech streaming default to SSE (#4679)

Signed-off-by: syd520zy <529477025@qq.com>

benchmarks/fish-speech/bench_speaker_cache.py

@@ -201,6 +201,7 @@ async def main():
             "input": prompt,
             "voice": "default",
             "stream": True,
+            "stream_format": "audio",
             "response_format": "pcm",
             "ref_audio": ref_audio_b64,
             "ref_text": args.ref_text,
@@ -240,6 +241,7 @@ async def main():
             "input": prompt,
             "voice": args.voice_name,
             "stream": True,
+            "stream_format": "audio",
             "response_format": "pcm",
             "ref_text": args.ref_text,
             "max_new_tokens": 2048,

docs/serving/speech_api.md

@@ -122,8 +122,8 @@ Content-Type: application/json
 | `max_new_tokens` | integer | 2048 | Maximum tokens to generate |
 | `initial_codec_chunk_frames` | integer | null | Per-request initial chunk size override for TTFA tuning. When null, IC is computed dynamically based on server load. |
 | `non_streaming_mode` | bool | null | Qwen3-TTS prompt construction mode override. Does not affect HTTP response streaming or async-chunk pipelining. When null, Qwen3-TTS uses model defaults: Base=false, CustomVoice/VoiceDesign=true. |
-| `stream_format` | string | null | Streaming output format. `"audio"` streams raw audio bytes as they are decoded; `"sse"` streams OpenAI `speech.audio.*` Server-Sent Events. Omit (null) for a single non-streaming binary response. See [Response Format](#response-format). |
-| `stream` | bool | false | Legacy streaming switch, equivalent to `stream_format="audio"`. Requires `response_format="pcm"` or `"wav"`. Speed adjustment is not supported when streaming. |
+| `stream` | bool | false | When true, stream OpenAI `speech.audio.*` SSE events (requires `response_format="pcm"` or `"wav"`). For raw PCM/WAV byte streaming, set `stream_format="audio"`. |
+| `stream_format` | string | null | Streaming output format. `"audio"` streams raw audio bytes as they are decoded; `"sse"` streams OpenAI `speech.audio.*` Server-Sent Events. If omitted, `stream=true` selects SSE and `stream=false` remains non-streaming. See [Response Format](#response-format). |
 
 **Supported languages:** Only applicable to Qwen3-TTS. Derived from the model configuration (`talker_config.codec_language_id` in the checkpoint's `config.json`), plus `Auto`, which is always accepted. Official Qwen3-TTS checkpoints support: Auto, Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, Italian.
 
@@ -143,18 +143,15 @@ The response shape depends on the streaming parameters:
 complete clip as binary audio data with an appropriate `Content-Type` header (e.g.
 `audio/wav`). The raw-bytes body has no JSON carrier, so no `usage` is reported.
 
-**Raw audio stream** (`stream=true` or `stream_format="audio"`). Streams raw
-audio bytes (PCM or WAV) as they are decoded.
+**Raw audio stream** (`stream_format="audio"`). Streams raw audio bytes (PCM or
+WAV) as they are decoded.
 
 Both streaming modes (`stream_format="audio"` and `"sse"`) require
 `response_format="pcm"` or `"wav"`, and `speed` must be `1.0` (or omitted).
 
-**SSE stream** (`stream_format="sse"`). Streams [OpenAI `speech.audio.*` Server-Sent
-Events](https://platform.openai.com/docs/api-reference/audio-streaming). Set
-`stream_format="sse"` and leave the legacy `stream` unset: `stream=true` takes
-precedence and forces the raw-audio path, so the response would be audio bytes with
-no `speech.audio.*` events (and no terminal `usage`). Each event has an `event:`
-line and a JSON `data:` line:
+**SSE stream** (`stream=true` or `stream_format="sse"`). Streams [OpenAI
+`speech.audio.*` Server-Sent Events](https://platform.openai.com/docs/api-reference/audio-streaming).
+Each event has an `event:` line and a JSON `data:` line:
 
 - `speech.audio.delta` — a base64 audio chunk:
 

docs/user_guide/examples/online_serving/text_to_speech.md

@@ -79,7 +79,7 @@ response = client.audio.speech.create(
 response.stream_to_file("output.wav")
 ```
 
-Streaming PCM output (where supported) — set `stream=true` with `response_format="pcm"`:
+Streaming PCM output (where supported) — set `stream=true`, `stream_format="audio"`, and `response_format="pcm"`:
 
 ```bash
 curl -X POST http://localhost:8091/v1/audio/speech \
@@ -88,6 +88,7 @@ curl -X POST http://localhost:8091/v1/audio/speech \
         "input": "Hello, how are you?",
         "voice": "default",
         "stream": true,
+        "stream_format": "audio",
         "response_format": "pcm"
     }' --no-buffer | play -t raw -r 24000 -e signed -b 16 -c 1 -
 ```
@@ -414,10 +415,11 @@ curl -X POST http://localhost:8091/v1/audio/speech \
         "voice": "vivian",
         "language": "English",
         "stream": true,
+        "stream_format": "audio",
         "response_format": "pcm"
     }' --no-buffer | play -t raw -r 24000 -e signed -b 16 -c 1 -
 ```
-Streaming requires `response_format="pcm"` and `async_chunk: true` on the stage config (default in `qwen3_tts.yaml`). `speed` is not supported when streaming.
+Raw PCM streaming requires `stream_format="audio"`, `response_format="pcm"`, and `async_chunk: true` on the stage config (default in `qwen3_tts.yaml`). `speed` is not supported when streaming.
 
 ### Streaming WebSocket
 The `/v1/audio/speech/stream` endpoint accepts text incrementally, splits it at sentence boundaries, and emits one PCM stream per sentence:
@@ -496,7 +498,7 @@ After startup, `/v1/audio/voices` lists `alice`, and `/v1/audio/speech` can use
 ```bash
 python voxcpm2/gradio_demo.py
 ```
-Uses an AudioWorklet-based player adapted from the Qwen3-TTS demo for gap-free playback. Audio is streamed from the OpenAI Speech endpoint with `stream=true`.
+Uses an AudioWorklet-based player adapted from the Qwen3-TTS demo for gap-free playback. Raw PCM audio is streamed from the OpenAI Speech endpoint with `stream=true` and `stream_format="audio"`.
 
 ---
 

examples/online_serving/text_to_speech/README.md

@@ -79,7 +79,7 @@ response = client.audio.speech.create(
 response.stream_to_file("output.wav")
 ```
 
-Streaming PCM output (where supported) — set `stream=true` with `response_format="pcm"`:
+Streaming PCM output (where supported) — set `stream=true`, `stream_format="audio"`, and `response_format="pcm"`:
 
 ```bash
 curl -X POST http://localhost:8091/v1/audio/speech \
@@ -88,6 +88,7 @@ curl -X POST http://localhost:8091/v1/audio/speech \
         "input": "Hello, how are you?",
         "voice": "default",
         "stream": true,
+        "stream_format": "audio",
         "response_format": "pcm"
     }' --no-buffer | play -t raw -r 24000 -e signed -b 16 -c 1 -
 ```
@@ -363,6 +364,7 @@ curl -X POST http://localhost:8091/v1/audio/speech \
         \"input\": \"Hello, streaming output from MOSS-TTS-Nano.\",
         \"ref_audio\": \"data:audio/wav;base64,${REF_AUDIO}\",
         \"stream\": true,
+        \"stream_format\": \"audio\",
         \"response_format\": \"pcm\"
     }" --no-buffer | play -t raw -r 48000 -e signed -b 16 -c 1 -
 ```
@@ -528,10 +530,11 @@ curl -X POST http://localhost:8091/v1/audio/speech \
         "voice": "vivian",
         "language": "English",
         "stream": true,
+        "stream_format": "audio",
         "response_format": "pcm"
     }' --no-buffer | play -t raw -r 24000 -e signed -b 16 -c 1 -
 ```
-Streaming requires `response_format="pcm"` and `async_chunk: true` on the stage config (default in `qwen3_tts.yaml`). `speed` is not supported when streaming.
+Raw PCM streaming requires `stream_format="audio"`, `response_format="pcm"`, and `async_chunk: true` on the stage config (default in `qwen3_tts.yaml`). `speed` is not supported when streaming.
 
 ### Streaming WebSocket
 The `/v1/audio/speech/stream` endpoint accepts text incrementally, splits it at sentence boundaries, and emits one PCM stream per sentence:
@@ -632,7 +635,7 @@ After startup, `/v1/audio/voices` lists `alice`, and `/v1/audio/speech` can use
 ```bash
 python voxcpm2/gradio_demo.py
 ```
-Uses an AudioWorklet-based player adapted from the Qwen3-TTS demo for gap-free playback. Audio is streamed from the OpenAI Speech endpoint with `stream=true`.
+Uses an AudioWorklet-based player adapted from the Qwen3-TTS demo for gap-free playback. Raw PCM audio is streamed from the OpenAI Speech endpoint with `stream=true` and `stream_format="audio"`.
 
 ---
 

examples/online_serving/text_to_speech/cosyvoice3/speech_client.py

@@ -60,6 +60,7 @@ def run_tts(args) -> None:
 
     if args.stream:
         payload["stream"] = True
+        payload["stream_format"] = "audio"
         payload["response_format"] = "pcm"
 
     print(f"Model: {args.model}")

examples/online_serving/text_to_speech/fish_speech/gradio_demo.py

@@ -66,6 +66,8 @@ def build_payload(
         "response_format": "pcm" if stream else response_format,
         "stream": stream,
     }
+    if stream:
+        payload["stream_format"] = "audio"
 
     # Voice cloning: ref_audio takes priority over URL.
     ref_url_stripped = ref_audio_url.strip() if ref_audio_url else ""

examples/online_serving/text_to_speech/fish_speech/speech_client.py

@@ -53,6 +53,7 @@ def run_tts(args) -> None:
 
     if args.stream:
         payload["stream"] = True
+        payload["stream_format"] = "audio"
         payload["response_format"] = "pcm"
 
     print(f"Model: {args.model}")

examples/online_serving/text_to_speech/glm_tts/gradio_demo.py

@@ -64,6 +64,8 @@ def build_payload(
         "stream": stream,
         "max_new_tokens": max_new_tokens,
     }
+    if stream:
+        payload["stream_format"] = "audio"
 
     # Voice cloning: ref_audio takes priority over URL.
     ref_url_stripped = ref_audio_url.strip() if ref_audio_url else ""

examples/online_serving/text_to_speech/glm_tts/openai_speech_client.py

@@ -40,6 +40,9 @@ def run_tts_generation(args) -> None:
         "ref_audio": args.ref_audio,
         "ref_text": args.ref_text,
     }
+    if args.stream:
+        payload["stream_format"] = "audio"
+        payload["response_format"] = "pcm"
     if args.max_new_tokens:
         payload["max_new_tokens"] = args.max_new_tokens
 

examples/online_serving/text_to_speech/ming_tts/openai_speech_client.py

@@ -123,6 +123,7 @@ def run_tts(args) -> None:
         payload["max_new_tokens"] = args.max_new_tokens
     if args.stream:
         payload["stream"] = True
+        payload["stream_format"] = "audio"
         payload["response_format"] = "pcm"
 
     api_url = f"{args.api_base}/v1/audio/speech"

examples/online_serving/text_to_speech/ming_tts/run_curl.sh

@@ -65,6 +65,7 @@ PY
             \"model\": \"${MODEL}\",
             \"input\": \"${TEXT}\",
             \"stream\": true,
+            \"stream_format\": \"audio\",
             \"response_format\": \"pcm\"
         }" "$STREAM_OUTPUT"
         ;;

examples/online_serving/text_to_speech/moss_tts_nano/gradio_demo.py

@@ -335,6 +335,7 @@ def build_payload(
         "ref_audio": encode_audio_to_base64(ref_audio),
         "response_format": "pcm" if stream else response_format,
         "stream": stream,
+        **({"stream_format": "audio"} if stream else {}),
     }
 
 

examples/online_serving/text_to_speech/qwen3_tts/tts_common.py

@@ -93,6 +93,8 @@ def build_payload(
         "response_format": "pcm" if stream else response_format,
         "stream": stream,
     }
+    if stream:
+        payload["stream_format"] = "audio"
     if not stream:
         payload["speed"] = speed
 

examples/online_serving/text_to_speech/voxcpm2/gradio_demo.py

@@ -515,6 +515,8 @@ def create_app(api_base: str):
                 "response_format": "pcm" if stream_enabled else "wav",
                 "stream": stream_enabled,
             }
+            if stream_enabled:
+                payload["stream_format"] = "audio"
 
             # Reference audio for cloning
             ref_url_s = ref_url.strip() if ref_url else ""

recipes/OpenBMB/VoxCPM2.md

@@ -149,6 +149,7 @@ curl -X POST http://127.0.0.1:8000/v1/audio/speech \
         "input": "Streaming PCM test on a single RTX 4090 with VoxCPM2.",
         "voice": "default",
         "stream": true,
+        "stream_format": "audio",
         "response_format": "pcm"
     }' --no-buffer | play -t raw -r 48000 -e signed -b 16 -c 1 -
 ```

recipes/OpenMOSS/MOSS-TTS.md

@@ -104,7 +104,8 @@ curl -X POST http://localhost:8091/v1/audio/speech \
         "voice": "default",
         "ref_audio": "https://raw.githubusercontent.com/OpenMOSS/MOSS-TTS/main/assets/audio/zh_1.wav",
         "response_format": "wav",
-        "stream": true
+        "stream": true,
+        "stream_format": "audio"
     }' --output output.wav
 ```
 

recipes/Qwen/Qwen3-TTS.md

@@ -148,6 +148,7 @@ curl -X POST http://localhost:8091/v1/audio/speech \
         "voice": "vivian",
         "language": "English",
         "stream": true,
+        "stream_format": "audio",
         "response_format": "pcm"
     }' --no-buffer | play -t raw -r 24000 -e signed -b 16 -c 1 -
 ```
@@ -163,7 +164,7 @@ python examples/offline_inference/text_to_speech/qwen3_tts/end2end.py --query-ty
 
 - Memory usage: The deploy config allocates `gpu_memory_utilization: 0.3` per stage (talker + code2wav share a single GPU). For the 0.6B variants or constrained GPUs, adjust via `--gpu-memory-utilization`.
 - Key flags: `--omni` is required. `--deploy-config` points to the bundled two-stage pipeline config.
-- Async chunking: Enabled by default in `qwen3_tts.yaml` for streaming-friendly first-audio latency. Streaming requires `stream=true` with `response_format="pcm"`.
+- Async chunking: Enabled by default in `qwen3_tts.yaml` for streaming-friendly first-audio latency. Raw audio streaming requires `stream=true`, `stream_format="audio"`, and `response_format="pcm"`.
 - Task/model matching: Each task type requires its matching model checkpoint. Using a CustomVoice model for a Base (voice clone) request will fail.
 - Known limitations: The server serves one model variant at a time. To switch task types (e.g., CustomVoice to Base), restart the server with the corresponding model.
 

recipes/mistralai/Voxtral-TTS.md

@@ -80,6 +80,7 @@ curl -X POST http://127.0.0.1:8091/v1/audio/speech \
     "voice": "casual_female",
     "language": "English",
     "stream": true,
+    "stream_format": "audio",
     "response_format": "pcm"
   }' \
   --output voxtral_stream.pcm

tests/e2e/online_serving/test_cosyvoice3_tts_expansion.py

@@ -104,6 +104,7 @@ def test_voice_clone_zh_002(omni_server, openai_client) -> None:
         "model": omni_server.model,
         "input": get_prompt("zh"),
         "stream": True,
+        "stream_format": "audio",
         "response_format": "wav",
         "ref_audio": REF_AUDIO_URL,
         "ref_text": REF_TEXT,

tests/e2e/online_serving/test_glm_tts_expansion.py

@@ -97,6 +97,7 @@ def test_voice_clone_zh_async_chunk(omni_server, openai_client) -> None:
         "model": omni_server.model,
         "input": "今天天气真不错，适合出去散散步。",
         "stream": True,
+        "stream_format": "audio",
         "response_format": "wav",
         "ref_audio": REF_AUDIO_URL,
         "ref_text": REF_TEXT,

tests/e2e/online_serving/test_higgs_audio_v2_expansion.py

@@ -98,6 +98,7 @@ class TestHiggsAudioV2OnlineHappyPath:
                 "model": omni_server.model,
                 "input": "Streaming the quick brown fox over the lazy dog.",
                 "stream": True,
+                "stream_format": "audio",
                 "response_format": "pcm",
                 "timeout": DEFAULT_SPEECH_TIMEOUT_S,
                 "min_audio_bytes": _MIN_AUDIO_BYTES,
@@ -270,6 +271,7 @@ class TestHiggsAudioV2OnlineVoiceClone:
                 "ref_audio": _REF_AUDIO_URL,
                 "ref_text": _REF_TEXT,
                 "stream": True,
+                "stream_format": "audio",
                 "response_format": "pcm",
                 "timeout": DEFAULT_SPEECH_TIMEOUT_S,
                 "min_audio_bytes": _MIN_AUDIO_BYTES,

tests/e2e/online_serving/test_higgs_audio_v3.py

@@ -135,6 +135,7 @@ class TestHiggsAudioV3OnlineHappyPath:
                 "model": omni_server.model,
                 "input": "The quick brown fox jumps over the lazy dog.",
                 "stream": True,
+                "stream_format": "audio",
                 "response_format": "pcm",
                 "timeout": DEFAULT_SPEECH_TIMEOUT_S,
                 "min_hnr_db": 0.0,
@@ -160,6 +161,7 @@ class TestHiggsAudioV3OnlineHappyPath:
                 "model": omni_server.model,
                 "input": "She sells seashells by the seashore.",
                 "stream": True,
+                "stream_format": "audio",
                 "response_format": "pcm",
                 "timeout": DEFAULT_SPEECH_TIMEOUT_S,
                 "min_hnr_db": -2.0,

tests/e2e/online_serving/test_indextts2.py

@@ -75,6 +75,7 @@ def test_streaming(omni_server, openai_client) -> None:
         "model": omni_server.model,
         "input": "Hello, this is a voice cloning demo.",
         "stream": True,
+        "stream_format": "audio",
         "response_format": "pcm",
         "ref_audio": REF_AUDIO_URL,
         "min_audio_bytes": 1024,

tests/e2e/online_serving/test_ming_tts_expansion.py

@@ -90,6 +90,7 @@ def test_text_to_audio_streaming_001(omni_server, openai_client) -> None:
         "input": get_prompt("zh_short"),
         "voice": "灵小甄",
         "stream": True,
+        "stream_format": "audio",
         "response_format": "wav",
         "timeout": 300.0,
     }

tests/e2e/online_serving/test_moss_tts_expansion.py

@@ -137,6 +137,7 @@ def test_text_to_audio_002_streaming(omni_server, openai_client, ref_audio_data_
         "model": omni_server.model,
         "input": get_prompt(),
         "stream": True,
+        "stream_format": "audio",
         "response_format": "pcm",
         "ref_audio": ref_audio_data_url,
         "min_hnr_db": -5.0,

tests/e2e/online_serving/test_moss_tts_nano_expansion.py

@@ -129,6 +129,7 @@ def test_text_to_audio_002(omni_server, openai_client, ref_audio_data_url) -> No
         "model": omni_server.model,
         "input": get_prompt(),
         "stream": True,
+        "stream_format": "audio",
         "response_format": "pcm",
         "ref_audio": ref_audio_data_url,
         "min_hnr_db": -5.0,

tests/e2e/online_serving/test_qwen3_tts_base.py

@@ -100,6 +100,7 @@ def test_text_to_audio_002(omni_server, openai_client) -> None:
         "model": omni_server.model,
         "input": get_prompt(),
         "stream": True,
+        "stream_format": "audio",
         "timeout": DEFAULT_AUDIO_SPEECH_TIMEOUT_S,
         "response_format": "wav",
         "task_type": "Base",

tests/e2e/online_serving/test_qwen3_tts_base_expansion.py

@@ -80,6 +80,7 @@ def test_voice_clone_streaming_001(omni_server, openai_client) -> None:
         "model": omni_server.model,
         "input": get_prompt(),
         "stream": True,
+        "stream_format": "audio",
         "response_format": "wav",
         "task_type": "Base",
         "voice": "clone",

tests/e2e/online_serving/test_qwen3_tts_customvoice.py

@@ -95,6 +95,7 @@ def test_text_to_audio_002(omni_server, openai_client) -> None:
         "model": omni_server.model,
         "input": get_prompt(),
         "stream": True,
+        "stream_format": "audio",
         "response_format": "wav",
         "task_type": "CustomVoice",
         "voice": "vivian",

tests/e2e/online_serving/test_qwen3_tts_customvoice_expansion.py

@@ -75,6 +75,7 @@ def test_voice_001(omni_server, openai_client) -> None:
         "model": omni_server.model,
         "input": get_prompt(),
         "stream": True,
+        "stream_format": "audio",
         "response_format": "wav",
         "task_type": "CustomVoice",
         "voice": "uncle_fu",

tests/e2e/online_serving/test_voxcpm2_tts.py

@@ -89,6 +89,7 @@ def test_text_to_audio_002(omni_server, openai_client) -> None:
         "model": omni_server.model,
         "input": get_prompt(),
         "stream": True,
+        "stream_format": "audio",
         "timeout": DEFAULT_AUDIO_SPEECH_TIMEOUT_S,
         "response_format": "wav",
         "voice": "default",

tests/e2e/online_serving/test_voxcpm2_tts_expansion.py

@@ -62,6 +62,7 @@ def test_voice_clone_streaming_001(omni_server, openai_client) -> None:
         "model": omni_server.model,
         "input": get_prompt(),
         "stream": True,
+        "stream_format": "audio",
         "timeout": DEFAULT_AUDIO_SPEECH_TIMEOUT_S,
         "response_format": "wav",
         "voice": "default",

tests/e2e/online_serving/test_voxtral_tts_expansion.py

@@ -56,6 +56,7 @@ def test_speech_english_streaming(omni_server, openai_client) -> None:
             "voice": "casual_female",
             "language": "English",
             "stream": True,
+            "stream_format": "audio",
             "response_format": "pcm",
             "timeout": 120.0,
             "min_audio_bytes": _MIN_AUDIO_BYTES_BASIC,

tests/entrypoints/openai_api/test_qwen3_tts_speaker_embedding.py

@@ -141,6 +141,7 @@ class TestSpeakerEmbeddingBase:
             "x_vector_only_mode": True,
             "response_format": "pcm",
             "stream": True,
+            "stream_format": "audio",
             "max_new_tokens": MAX_NEW_TOKENS,
         }
         with httpx.Client(timeout=120.0) as client:

tests/entrypoints/openai_api/test_serving_speech.py

@@ -1849,17 +1849,22 @@ class TestStreamingProtocolValidation:
             OpenAICreateSpeechRequest(input="Hello", stream_format="audio", response_format="pcm", speed=2.0)
 
     def test_stream_valid(self):
-        """stream=True + response_format in ('pcm', 'wav') + speed=1.0 is accepted."""
+        """stream=True + response_format in ('pcm', 'wav') + speed=1.0 is accepted as SSE."""
         req = OpenAICreateSpeechRequest(input="Hello", stream=True, response_format="pcm")
         assert req.stream is True
+        assert req.is_sse_stream() is True
+        assert req.is_raw_audio_stream() is False
 
         req = OpenAICreateSpeechRequest(input="Hello", stream=True, response_format="wav")
         assert req.stream is True
+        assert req.is_sse_stream() is True
+        assert req.is_raw_audio_stream() is False
 
     def test_stream_format_audio_is_valid(self):
         req = OpenAICreateSpeechRequest(input="Hello", stream_format="audio", response_format="pcm")
         assert req.stream_format == "audio"
         assert req.is_raw_audio_stream() is True
+        assert req.is_sse_stream() is False
 
     def test_sse_stream_format_is_valid(self):
         """stream_format='sse' is accepted for /audio/speech."""
@@ -1867,6 +1872,7 @@ class TestStreamingProtocolValidation:
 
         assert req.stream_format == "sse"
         assert req.is_sse_stream() is True
+        assert req.is_raw_audio_stream() is False
 
 
 class TestStreamingResponse:
@@ -1939,13 +1945,24 @@ class TestStreamingResponse:
         app.add_api_route("/v1/audio/speech", speech_server.create_speech, methods=["POST"], response_model=None)
         return app
 
+    @staticmethod
+    def _assert_sse_audio_response(response, response_format: str = "pcm"):
+        assert response.status_code == 200
+        assert "text/event-stream" in response.headers["content-type"]
+        body = response.text
+        assert "event: speech.audio.delta" in body
+        assert "event: speech.audio.done" in body
+        data_line = next(line for line in body.splitlines() if line.startswith("data: "))
+        payload = json.loads(data_line.removeprefix("data: "))
+        assert payload["type"] == "speech.audio.delta"
+        assert payload["response_format"] == response_format
+        assert base64.b64decode(payload["audio"])
+
     def test_streaming(self, streaming_app):
-        """stream=True must return audio/pcm with non-empty body."""
+        """stream=True defaults to OpenAI speech.audio.* SSE events."""
         client = TestClient(streaming_app)
         response = client.post("/v1/audio/speech", json={"input": "Hello", "stream": True, "response_format": "pcm"})
-        assert response.status_code == 200
-        assert "audio/pcm" in response.headers["content-type"]
-        assert len(response.content) > 0
+        self._assert_sse_audio_response(response)
 
     def test_stream_format_audio_streaming(self, streaming_app):
         """stream_format=audio without stream=True returns raw audio/pcm chunks."""
@@ -1967,25 +1984,26 @@ class TestStreamingResponse:
             json={"input": "Hello", "stream_format": "sse", "response_format": "pcm"},
         )
 
-        assert response.status_code == 200
-        assert "text/event-stream" in response.headers["content-type"]
-        body = response.text
-        assert "event: speech.audio.delta" in body
-        assert "event: speech.audio.done" in body
-        data_line = next(line for line in body.splitlines() if line.startswith("data: "))
-        payload = json.loads(data_line.removeprefix("data: "))
-        assert payload["type"] == "speech.audio.delta"
-        assert payload["response_format"] == "pcm"
-        assert base64.b64decode(payload["audio"])
+        self._assert_sse_audio_response(response)
 
-    def test_stream_true_prefers_raw_audio_over_sse(self, streaming_app):
-        """stream=True keeps the existing raw audio stream behavior even with stream_format=sse."""
+    def test_stream_true_with_stream_format_sse_uses_sse(self, streaming_app):
+        """stream=True and stream_format=sse both select SSE streaming."""
         client = TestClient(streaming_app)
         response = client.post(
             "/v1/audio/speech",
             json={"input": "Hello", "stream": True, "stream_format": "sse", "response_format": "pcm"},
         )
 
+        self._assert_sse_audio_response(response)
+
+    def test_stream_format_audio_with_stream_true_opts_into_raw_audio(self, streaming_app):
+        """stream_format=audio remains an explicit raw audio opt-in."""
+        client = TestClient(streaming_app)
+        response = client.post(
+            "/v1/audio/speech",
+            json={"input": "Hello", "stream": True, "stream_format": "audio", "response_format": "pcm"},
+        )
+
         assert response.status_code == 200
         assert "audio/pcm" in response.headers["content-type"]
         assert "text/event-stream" not in response.headers["content-type"]
@@ -3021,7 +3039,10 @@ class TestWAVStreaming:
     def test_wav_streaming_success(self, wav_streaming_app):
         """Test WAV format streaming returns correct content type and includes WAV header."""
         client = TestClient(wav_streaming_app)
-        response = client.post("/v1/audio/speech", json={"input": "Hello", "stream": True, "response_format": "wav"})
+        response = client.post(
+            "/v1/audio/speech",
+            json={"input": "Hello", "stream": True, "stream_format": "audio", "response_format": "wav"},
+        )
 
         assert response.status_code == 200
         assert "audio/wav" in response.headers["content-type"]
@@ -3040,7 +3061,10 @@ class TestWAVStreaming:
 
         unsupported_formats = ["mp3"]
         for fmt in unsupported_formats:
-            response = client.post("/v1/audio/speech", json={"input": "Hello", "stream": True, "response_format": fmt})
+            response = client.post(
+                "/v1/audio/speech",
+                json={"input": "Hello", "stream": True, "stream_format": "audio", "response_format": fmt},
+            )
             assert response.status_code == 422
 
 
@@ -3336,6 +3360,23 @@ class TestTTSAsyncOffloading:
 
         mock_coerce.assert_called_once_with(qwen3_tts_server.engine_client.default_sampling_params_list, True)
 
+    def test_prepare_speech_generation_treats_stream_true_as_streaming(self, qwen3_tts_server, mocker: MockerFixture):
+        """stream=True should request delta-style multimodal outputs for SSE streaming."""
+        qwen3_tts_server._validate_tts_request = mocker.MagicMock(return_value=None)
+        qwen3_tts_server._build_tts_params = mocker.MagicMock(
+            return_value={"text": ["hello"], "task_type": ["CustomVoice"], "speaker": ["Vivian"]}
+        )
+        qwen3_tts_server._estimate_prompt_len_async = mocker.AsyncMock(return_value=512)
+        mock_coerce = mocker.patch(
+            "vllm_omni.entrypoints.openai.serving_speech.coerce_param_message_types",
+            return_value=qwen3_tts_server.engine_client.default_sampling_params_list,
+        )
+        request = OpenAICreateSpeechRequest(input="hello", stream=True, response_format="pcm")
+
+        asyncio.run(qwen3_tts_server._prepare_speech_generation(request))
+
+        mock_coerce.assert_called_once_with(qwen3_tts_server.engine_client.default_sampling_params_list, True)
+
     def test_prepare_speech_generation_treats_audio_as_streaming(self, qwen3_tts_server, mocker: MockerFixture):
         """stream_format=audio should request delta-style multimodal outputs."""
         qwen3_tts_server._validate_tts_request = mocker.MagicMock(return_value=None)

tests/helpers/runtime.py

@@ -1770,6 +1770,7 @@ class OpenAIClientHandler:
             "seed",
             "instructions",
             "speed",
+            "stream_format",
         ):
             if key in request_config:
                 extra_body[key] = request_config[key]

vllm_omni/benchmarks/patch/patch.py

@@ -1133,9 +1133,9 @@ async def async_request_openai_audio_speech(
 ) -> MixRequestFuncOutput:
     """Streaming request to /v1/audio/speech endpoint.
 
-    Sends ``stream=true`` with ``response_format=pcm`` so the server returns
-    raw PCM chunks as they are decoded. This allows measuring TTFP (time to
-    first audio packet) separately from E2EL.
+    Sends ``stream=true`` with ``stream_format=audio`` and ``response_format=pcm``
+    so the server returns raw PCM chunks as they are decoded. This allows measuring
+    TTFP (time to first audio packet) separately from E2EL.
     """
     api_url = request_func_input.api_url
     _validate_api_url(api_url, "OpenAI Audio Speech API", "audio/speech")
@@ -1144,12 +1144,14 @@ async def async_request_openai_audio_speech(
         "model": request_func_input.model_name if request_func_input.model_name else request_func_input.model,
         "input": request_func_input.prompt,
         "stream": True,
+        "stream_format": "audio",
         "response_format": "pcm",
     }
     _update_payload_common(payload, request_func_input)
     # Seed-TTS + WER: ``--extra-body`` may set stream=false / other formats; speech must stream PCM.
     if getattr(request_func_input, "seed_tts_row", False) and _seed_tts_capture_pcm_for_wer():
         payload["stream"] = True
+        payload["stream_format"] = "audio"
         payload["response_format"] = "pcm"
 
     headers = {
@@ -1219,7 +1221,7 @@ async def async_request_openai_audio_speech(
                     ct = response.headers.get("Content-Type", "")
                     logger.warning(
                         "Seed-TTS WER: HTTP 200 but no PCM bytes (Content-Type=%r, url=%s). "
-                        "Check stream=true and response_format=pcm on the server.",
+                        "Check stream=true, stream_format=audio, and response_format=pcm on the server.",
                         ct,
                         api_url,
                     )

vllm_omni/entrypoints/openai/protocol/audio.py

@@ -69,13 +69,15 @@ class OpenAICreateSpeechRequest(BaseModel):
         default=None,
         description=(
             "Streaming output format. 'audio' streams raw pcm/wav bytes; "
-            "'sse' streams OpenAI speech.audio.* SSE events. Omit for non-streaming."
+            "'sse' streams OpenAI speech.audio.* SSE events. If omitted, stream=true "
+            "selects SSE and stream=false remains non-streaming."
         ),
     )
     stream: bool = Field(
         default=False,
         description=(
-            "Legacy streaming switch; equivalent to stream_format='audio'. "
+            "Streaming switch; defaults to OpenAI speech.audio.* SSE events. "
+            "Set stream_format='audio' to opt into raw pcm/wav byte streaming. "
             "Requires response_format='pcm' or 'wav'. Speed adjustment is not supported when streaming."
         ),
     )
@@ -292,13 +294,13 @@ class OpenAICreateSpeechRequest(BaseModel):
         return self
 
     def is_raw_audio_stream(self) -> bool:
-        return self.stream or self.stream_format == "audio"
+        return self.stream_format == "audio"
 
     def is_sse_stream(self) -> bool:
-        return self.stream_format == "sse" and not self.is_raw_audio_stream()
+        return (self.stream or self.stream_format == "sse") and not self.is_raw_audio_stream()
 
     def is_streaming(self) -> bool:
-        return self.is_raw_audio_stream() or self.stream_format == "sse"
+        return self.is_raw_audio_stream() or self.is_sse_stream()
 
     @model_validator(mode="after")
     def validate_streaming_constraints(self) -> "OpenAICreateSpeechRequest":

vllm_omni/entrypoints/openai/serving_speech.py

@@ -4038,11 +4038,11 @@ class OmniOpenAIServingSpeech(OpenAIServing, AudioMixin):
         - ref_text: Transcript of reference audio (Base task)
         - x_vector_only_mode: Use speaker embedding only (Base task)
 
-        Streaming is supported via ``stream_format='audio'`` or the legacy
-        ``stream=True`` switch, with ``response_format='pcm'`` or ``'wav'``.
-        ``stream_format='sse'`` returns OpenAI ``speech.audio.*`` SSE events instead.
-        Each Code2Wav chunk is yielded as raw audio bytes as soon as it is decoded.
-        For WAV format, a header with placeholder size values is emitted first.
+        Streaming is supported via the ``stream=True`` switch or ``stream_format='sse'``,
+        which return OpenAI ``speech.audio.*`` SSE events. ``stream_format='audio'``
+        opts into raw audio streaming with ``response_format='pcm'`` or ``'wav'``.
+        Raw audio streaming yields each Code2Wav chunk as raw bytes as soon as it is
+        decoded. Raw WAV streaming emits a header with placeholder size values first.
         """
         if self._diffusion_mode:
             return await self._create_diffusion_speech(request)
@@ -4060,14 +4060,14 @@ class OmniOpenAIServingSpeech(OpenAIServing, AudioMixin):
             )
 
         try:
-            if request.is_raw_audio_stream():
-                if request.word_timestamps:
+            if request.is_streaming() and request.word_timestamps:
                 return self.create_error_response(
                     "word_timestamps=true is currently supported by the WebSocket "
                     "/v1/audio/speech/stream path. Use session.config with "
                     "stream_audio=true and response_format='pcm'."
                 )
 
+            if request.is_raw_audio_stream():
                 response_format, error = self._validate_speech_streaming_request(
                     request,
                     mode_label="Streaming",