Voice SDK: Updated ForceEndOfUtterance to support padding the timestamp

sdk/voice/pyproject.toml

@@ -11,7 +11,7 @@ authors = [{ name = "Speechmatics", email = "support@speechmatics.com" }]
 license = "MIT"
 requires-python = ">=3.9"
 dependencies = [
-    "speechmatics-rt>=0.5.3",
+    "speechmatics-rt>=1.0.0",
     "pydantic>=2.10.6,<3",
     "numpy>=1.26.4,<3"
 ]
@@ -42,7 +42,7 @@ keywords = [
 smart = [
     "certifi>=2025.10.5",
     "onnxruntime>=1.20.1,<2",
-    "transformers>=4.57.0,<5",
+    "transformers>=4.57.0,<6",
 ]
 dev = [
     "black",

sdk/voice/speechmatics/voice/_client.py

@@ -717,14 +717,14 @@ def update_diarization_config(self, config: SpeakerFocusConfig) -> None:
     # PUBLIC UTTERANCE / TURN MANAGEMENT
     # ============================================================================
 
-    def finalize(self, end_of_turn: bool = False) -> None:
+    def finalize(self, pad: float | None = None) -> None:
         """Finalize segments.
 
         This function will emit segments in the buffer without any further checks
         on the contents of the segments.
 
         Args:
-            end_of_turn: Whether to emit an end of turn message.
+            pad: the number of seconds to pad the timestamp for the FEOU (optional)
         """
 
         # Clear smart turn cutoff
@@ -738,8 +738,8 @@ async def emit() -> None:
             """Wait for EndOfUtterance if needed, then emit segments."""
 
             # Forced end of utterance message (only when no speaker is detected)
-            if self._config.end_of_turn_config.use_forced_eou:
-                await self._await_forced_eou()
+            if self._uses_forced_eou:
+                await self._await_forced_eou(pad=pad)
 
             # Check if the turn has changed
             if self._turn_handler.handler_id != _turn_id:
@@ -1667,7 +1667,7 @@ async def _smart_turn_prediction(
         # Return the prediction
         return prediction
 
-    async def _await_forced_eou(self, timeout: float = 1.0) -> None:
+    async def _await_forced_eou(self, timeout: float = 1.0, pad: float | None = None) -> None:
         """Await the forced end of utterance."""
 
         # Received EOU
@@ -1676,17 +1676,36 @@ async def _await_forced_eou(self, timeout: float = 1.0) -> None:
         # Add listener
         self.once(AgentServerMessageType.END_OF_UTTERANCE, lambda message: eou_received.set())
 
-        # Trigger EOU message
-        self._emit_diagnostic_message("ForceEndOfUtterance sent - waiting for EndOfUtterance")
-
         # Wait for EOU
         try:
             # Track the start time
             start_time = time.time()
             self._forced_eou_active = True
 
+            # Timings
+            audio_sent = self.audio_seconds_sent
+
+            # Padding precedence: explicit `pad` arg > SPEECHMATICS_FEOU_PAD env var > config
+            if pad is not None:
+                padding = pad
+            else:
+                env_pad = os.environ.get("SPEECHMATICS_FEOU_PAD")
+                if env_pad is not None:
+                    padding = float(env_pad)
+                else:
+                    padding = self._config.end_of_turn_config.forced_eou_padding
+
+            # Establish amount of time to wait for EOU
+            timestamp: float = max(audio_sent + padding, 0.0)
+
+            # Info
+            info = {"audio_sent": audio_sent, "padding": padding, "timestamp": timestamp}
+
             # Send the force EOU and wait for the response
-            await self.force_end_of_utterance()
+            await self.force_end_of_utterance(timestamp=timestamp)
+            self._emit_diagnostic_message(f"ForceEndOfUtterance sent - waiting for EndOfUtterance ({info})")
+
+            # Wait for the response
             await asyncio.wait_for(eou_received.wait(), timeout=timeout)
 
             # Record the latency

sdk/voice/speechmatics/voice/_models.py

@@ -372,7 +372,7 @@ class SpeechSegmentConfig(BaseModel):
 
     Parameters:
         add_trailing_eos: Add trailing end of sentence to segments. When enabled, segments are
-            emitted with missing trailing end of sentence added. Defaults to False.
+            emitted with missing trailing end of sentence added. Defaults to True.
 
         emit_sentences: Emit segments when a sentence has ended. A finalized segment is emitted
             as soon as a finalized end of sentence is detected. If a speaker continues to speak during
@@ -384,7 +384,7 @@ class SpeechSegmentConfig(BaseModel):
             Defaults to None.
     """
 
-    add_trailing_eos: bool = False
+    add_trailing_eos: bool = True
     emit_sentences: bool = True
     pause_mark: Optional[str] = None
 
@@ -411,6 +411,7 @@ class EndOfTurnConfig(BaseModel):
         min_end_of_turn_delay: Minimum end of turn delay.
         penalties: List of end of turn penalty items.
         use_forced_eou: Whether to use forced end of utterance detection.
+        forced_eou_padding: the padding to use when sending ForceEndOfUtterance with timestamp
     """
 
     base_multiplier: float = 1.0
@@ -439,6 +440,7 @@ class EndOfTurnConfig(BaseModel):
         ]
     )
     use_forced_eou: bool = False
+    forced_eou_padding: float = 0.2
 
 
 class VoiceActivityConfig(BaseModel):