Add support for emit_sentences configuration

examples/agent_stt/realtime_file/main.py

@@ -43,6 +43,11 @@ def parse_args() -> argparse.Namespace:
         "--turn-seconds", type=float, default=5.0, help="fake turn length, --turn-detection external only"
     )
     parser.add_argument("--no-partials", action="store_true")
+    parser.add_argument(
+        "--emit-sentences",
+        action="store_true",
+        help="one sentence per segment, split on end-of-sentence punctuation: 'Hi. How are you?' arrives as two segments",
+    )
     return parser.parse_args()
 
 
@@ -75,6 +80,7 @@ def build_client(args: argparse.Namespace, clock: Clock) -> AgentSttAsyncClient:
     transcription_config = TranscriptionConfig(
         language=args.language,
         enable_partials=not args.no_partials,
+        emit_sentences=args.emit_sentences or None,
     )
     turn_config = TurnConfig(turn_detection_mode=TurnDetectionMode(args.turn_detection))
 

sdk/agent_stt/README.md

@@ -127,6 +127,14 @@ point, so the transcriber never sees a name it has no notion of. The RT models (
 `standard`) are not Agent STT models and are not accepted here; the deprecated `operating_point`
 still passes through, and suppresses the `model` default so the two never arrive together.
 
+`emit_sentences` splits a segment after each end-of-sentence punctuation mark (such as `.`, `?`
+or `!`), so each segment carries one sentence: "Hi there. How are you?" arrives as two segments
+rather than one. Off unless you ask for it, and fixed for the life of the session:
+
+```python
+transcription_config = TranscriptionConfig(language="en", emit_sentences=True)
+```
+
 Engine silence-based end of utterance is not offered here. A turn ends either because the
 service's VAD said so, or because you called `finalize()`.
 

sdk/agent_stt/speechmatics/agent_stt/_models.py

@@ -182,16 +182,22 @@ class TranscriptionConfig(RTTranscriptionConfig):
     Attributes:
         additional_vocab: Words to bias the engine towards, as `AdditionalVocabEntry` objects
             or raw dicts.
-        model: Agent STT model, defaulting to `DEFAULT_MODEL`. The proxy in front of the
-            service resolves the name to the engine's operating point, so a model the
-            transcriber has no notion of still routes correctly.
+        model: Agent STT model to transcribe with. Defaults to `DEFAULT_MODEL`.
+        emit_sentences: Split a segment after each end-of-sentence punctuation mark (such as
+            `.`, `?` or `!`), so each segment carries a single sentence. For example,
+            "Hi there. How are you?" arrives as two segments rather than one. Fixed for the
+            life of the session. Defaults to False service-side.
 
     Examples:
         >>> transcription_config = TranscriptionConfig(language="en", enable_partials=True)
+
+        One sentence per segment:
+            >>> transcription_config = TranscriptionConfig(language="en", emit_sentences=True)
     """
 
     model: Model = _UNSET
     additional_vocab: Optional[list[Union[AdditionalVocabEntry, dict[str, Any]]]] = None
+    emit_sentences: Optional[bool] = None
 
     def __post_init__(self) -> None:
         if self.model is not _UNSET and self.operating_point is not None:  # type: ignore[unreachable]

tests/agent_stt/test_client.py

@@ -268,6 +268,31 @@ async def test_start_session_turn_config_wins_over_the_constructor(client, monke
     assert transport.messages[0]["turn_config"] == {"turn_detection_mode": "vad"}
 
 
+@pytest.mark.asyncio
+async def test_emit_sentences_reaches_start_recognition(client, monkeypatch):
+    """The whole config dict is sent, but assert it rather than trust it: the flag is the session."""
+    transport = StubTransport()
+    client._transport = transport
+    monkeypatch.setattr(client, "_ws_connect", _noop)
+    monkeypatch.setattr(client, "_wait_recognition_started", _noop)
+
+    await client.start_session(transcription_config=TranscriptionConfig(language="en", emit_sentences=True))
+
+    assert transport.messages[0]["transcription_config"]["emit_sentences"] is True
+
+
+@pytest.mark.asyncio
+async def test_emit_sentences_absent_when_not_asked_for(client, monkeypatch):
+    transport = StubTransport()
+    client._transport = transport
+    monkeypatch.setattr(client, "_ws_connect", _noop)
+    monkeypatch.setattr(client, "_wait_recognition_started", _noop)
+
+    await client.start_session(transcription_config=TranscriptionConfig(language="en"))
+
+    assert "emit_sentences" not in transport.messages[0]["transcription_config"]
+
+
 def test_session_info_tracks_the_request_id_across_sessions(client):
     """RT mints a fresh request_id per session; session_info must not keep the first one."""
     start_session(client)

tests/agent_stt/test_config.py

@@ -11,9 +11,7 @@ def test_service_vad_is_the_default():
 
 
 def test_external_mode_selects_external_turn_detection():
-    assert TurnConfig(turn_detection_mode=TurnDetectionMode.EXTERNAL).to_dict() == {
-        "turn_detection_mode": "external"
-    }
+    assert TurnConfig(turn_detection_mode=TurnDetectionMode.EXTERNAL).to_dict() == {"turn_detection_mode": "external"}
 
 
 def test_turn_detection_is_not_in_the_transcription_config():
@@ -48,6 +46,16 @@ def test_rt_fields_still_work():
     assert result["max_delay"] == 1.5
 
 
+def test_emit_sentences_omitted_by_default():
+    """Left out so the service applies its own default of false."""
+    assert "emit_sentences" not in TranscriptionConfig(language="en").to_dict()
+
+
+@pytest.mark.parametrize("value", [True, False])
+def test_emit_sentences_sent_when_set(value):
+    assert TranscriptionConfig(language="en", emit_sentences=value).to_dict()["emit_sentences"] is value
+
+
 def test_model_and_operating_point_conflict():
     with pytest.raises(ValueError):
         TranscriptionConfig(model=Model.LINDEN_1, operating_point="enhanced")