dv-pipecat-ai 0.0.93.dev149__tar.gz → 0.0.93.dev151__tar.gz

{dv_pipecat_ai-0.0.93.dev149/src/dv_pipecat_ai.egg-info → dv_pipecat_ai-0.0.93.dev151}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: dv-pipecat-ai
-Version: 0.0.93.dev149
+Version: 0.0.93.dev151
 Summary: An open source framework for voice (and multimodal) assistants
 License-Expression: BSD-2-Clause
 Project-URL: Source, https://github.com/pipecat-ai/pipecat

{dv_pipecat_ai-0.0.93.dev149 → dv_pipecat_ai-0.0.93.dev151/src/dv_pipecat_ai.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: dv-pipecat-ai
-Version: 0.0.93.dev149
+Version: 0.0.93.dev151
 Summary: An open source framework for voice (and multimodal) assistants
 License-Expression: BSD-2-Clause
 Project-URL: Source, https://github.com/pipecat-ai/pipecat

dv_pipecat_ai-0.0.93.dev151/src/dv_pipecat_ai.egg-info/scm_version.json

@@ -0,0 +1,8 @@
+{
+  "tag": "0.0.92",
+  "distance": 151,
+  "node": "g68d849a09a9b507fb5f2ac472e81b90338d78f10",
+  "dirty": false,
+  "branch": "dv-stage",
+  "node_date": "2026-09-13"
+}

{dv_pipecat_ai-0.0.93.dev149 → dv_pipecat_ai-0.0.93.dev151}/src/pipecat/services/soniox/stt.py

@@ -92,6 +92,9 @@ class SonioxInputParams(BaseModel):
         enable_speaker_diarization: Whether to enable speaker diarization. Tokens are annotated with speaker IDs.
         enable_language_identification: Whether to enable language identification. Tokens are annotated with language IDs.
         client_reference_id: Client reference ID to use for transcription.
+        endpoint_sensitivity: How readily the model commits an endpoint (-1.0 to 1.0).
+        endpoint_latency_adjustment_level: Trade accuracy for lower endpoint latency (0 to 3).
+        max_endpoint_delay_ms: Hard ceiling on the endpoint wait after speech ends (500 to 3000).
     """
 
     model: str = "stt-rt-preview"
@@ -107,6 +110,15 @@ class SonioxInputParams(BaseModel):
 
     client_reference_id: Optional[str] = None
 
+    # Semantic-endpointing tuning. Only has an effect when the service runs with
+    # `vad_force_turn_endpoint=False` -- i.e. Soniox, not a silence VAD, decides
+    # where the turn ends. None leaves Soniox on its own default (sensitivity 0.0,
+    # latency adjustment 0, max delay 2000ms).
+    # See https://soniox.com/docs/stt/rt/endpoint-detection
+    endpoint_sensitivity: Optional[float] = None  # -1.0 .. 1.0, higher = commits sooner
+    endpoint_latency_adjustment_level: Optional[int] = None  # 0 .. 3, higher = lower latency
+    max_endpoint_delay_ms: Optional[int] = None  # 500 .. 3000
+
 
 # Connecting must never block StartFrame propagation. `AIService.process_frame`
 # awaits `start()` BEFORE `STTService.process_frame` pushes the StartFrame
@@ -333,6 +345,17 @@ class SonioxSTTService(STTService):
             "client_reference_id": self._params.client_reference_id,
         }
 
+        # Endpoint tuning is omitted entirely when unset so Soniox applies its own
+        # defaults; only meaningful while `enable_endpoint_detection` is on.
+        for key in (
+            "endpoint_sensitivity",
+            "endpoint_latency_adjustment_level",
+            "max_endpoint_delay_ms",
+        ):
+            value = getattr(self._params, key)
+            if value is not None:
+                config[key] = value
+
         # Send the configuration message.
         await self._websocket.send(json.dumps(config))
 
@@ -506,6 +529,40 @@ class SonioxSTTService(STTService):
                         if is_end_token(token):
                             # Found an endpoint, tokens until here will be sent as transcript,
                             # the rest will be sent as interim tokens (even final tokens).
+                            #
+                            # Logged explicitly because the decision is otherwise
+                            # invisible downstream: the only trace is the
+                            # TranscriptionFrame's wall-clock timestamp, which
+                            # conflates "Soniox decided" with "we pushed it".
+                            #
+                            # `<end>` is the model's own semantic call; `<fin>` is
+                            # only the ack of a `finalize` WE sent. They are reported
+                            # separately on purpose -- scoring endpoint quality over
+                            # `<fin>` would just be scoring our own VAD.
+                            #
+                            # Offsets are Soniox's audio positions, not wall clock,
+                            # so `lag_ms` (how long after speech stopped the endpoint
+                            # was placed) is immune to network and event-loop skew and
+                            # lines up directly with a call recording.
+                            buffered = self._final_transcription_buffer
+                            last = buffered[-1] if buffered else None
+                            speech_end_ms = (
+                                last.get("start_ms", 0) + last.get("duration_ms", 0)
+                                if last
+                                else None
+                            )
+                            at_ms = token.get("start_ms")
+                            lag_ms = (
+                                at_ms - speech_end_ms
+                                if at_ms is not None and speech_end_ms is not None
+                                else None
+                            )
+                            source = "model" if token["text"] == END_TOKEN else "finalize"
+                            logger.debug(
+                                f"Soniox endpoint: source={source}, at_ms={at_ms}, "
+                                f"speech_end_ms={speech_end_ms}, lag_ms={lag_ms}, "
+                                f"tokens={len(buffered)}"
+                            )
                             await send_endpoint_transcript()
                         else:
                             self._final_transcription_buffer.append(token)

dv_pipecat_ai-0.0.93.dev149/src/dv_pipecat_ai.egg-info/scm_version.json

@@ -1,8 +0,0 @@
-{
-  "tag": "0.0.92",
-  "distance": 149,
-  "node": "g2e17180069f8315a22c7718c9b7291a97bb74d50",
-  "dirty": false,
-  "branch": "dv-stage",
-  "node_date": "2026-09-09"
-}