docs(assemblyai): add vad_threshold and remove legacy parameters

- Add vad_threshold parameter documentation for U3 Pro
- Remove formatted_finals (v2 API legacy parameter)
- Remove word_finalization_max_wait_time (v2 API legacy parameter)
- Clarify format_turns only applies to Universal-Streaming models
- Add VAD threshold alignment usage example and notes

This corresponds to the code changes in pipecat-ai/pipecat#3927

Author: zkleb-aai

server/services/stt/assemblyai.mdx

@@ -125,8 +125,6 @@ Connection-level parameters passed via the `connection_params` constructor argum
 | ---------------------------------------- | ----------- | -------------- | --------------------------------------------------------------------------------------------------------------------------------------- |
 | `sample_rate`                            | `int`       | `16000`        | Audio sample rate in Hz.                                                                                                                |
 | `encoding`                               | `Literal`   | `"pcm_s16le"`  | Audio encoding format. Options: `"pcm_s16le"`, `"pcm_mulaw"`.                                                                           |
-| `formatted_finals`                       | `bool`      | `True`         | Whether to enable transcript formatting.                                                                                                |
-| `word_finalization_max_wait_time`        | `int`       | `None`         | Maximum time to wait for word finalization in milliseconds.                                                                             |
 | `end_of_turn_confidence_threshold`       | `float`     | `None`         | Confidence threshold for end-of-turn detection.                                                                                         |
 | `min_turn_silence`                       | `int`       | `None`         | Minimum silence duration (ms) when confident about end-of-turn.                                                                         |
 | `min_end_of_turn_silence_when_confident` | `int`       | `None`         | **DEPRECATED**. Use `min_turn_silence` instead. Will be removed in a future version.                                                    |
@@ -135,18 +133,23 @@ Connection-level parameters passed via the `connection_params` constructor argum
 | `prompt`                                 | `str`       | `None`         | Optional text prompt to guide transcription. Only used when `speech_model` is `"u3-rt-pro"`. Cannot be used with `keyterms_prompt`.    |
 | `speech_model`                           | `Literal`   | `"u3-rt-pro"`  | Speech model. Options: `"universal-streaming-english"`, `"universal-streaming-multilingual"`, `"u3-rt-pro"`.                            |
 | `language_detection`                     | `bool`      | `None`         | Enable automatic language detection. Only applicable to `universal-streaming-multilingual`. Turn messages include language information. |
-| `format_turns`                           | `bool`      | `True`         | Whether to format transcript turns.                                                                                                     |
+| `format_turns`                           | `bool`      | `True`         | Whether to format transcript turns. Only applicable to `universal-streaming-english` and `universal-streaming-multilingual` models. For `u3-rt-pro`, formatting is automatic and built-in. |
 | `speaker_labels`                         | `bool`      | `None`         | Enable speaker diarization. Final transcripts include a speaker field (e.g., "Speaker A", "Speaker B").                                 |
+| `vad_threshold`                          | `float`     | `None`         | Voice activity detection confidence threshold. Only applicable to `u3-rt-pro`. The confidence threshold (0.0 to 1.0) for classifying audio frames as silence. Frames with VAD confidence below this value are considered silent. Increase for noisy environments to reduce false speech detection. Defaults to 0.3 (API default). For best performance when using with external VAD (e.g., Silero), align this value with your VAD's activation threshold. Defaults to `None` (not sent). |
 
 ## Usage
 
 ### Basic Setup
 
 ```python
 from pipecat.services.assemblyai import AssemblyAISTTService
+from pipecat.services.assemblyai.models import AssemblyAIConnectionParams
 
 stt = AssemblyAISTTService(
     api_key=os.getenv("ASSEMBLYAI_API_KEY"),
+    connection_params=AssemblyAIConnectionParams(
+        speech_model="u3-rt-pro",
+    ),
 )
 ```
 
@@ -160,7 +163,6 @@ stt = AssemblyAISTTService(
     api_key=os.getenv("ASSEMBLYAI_API_KEY"),
     connection_params=AssemblyAIConnectionParams(
         sample_rate=16000,
-        formatted_finals=True,
         keyterms_prompt=["Pipecat", "AssemblyAI"],
         speech_model="u3-rt-pro",
     ),
@@ -206,9 +208,43 @@ stt = AssemblyAISTTService(
 )
 ```
 
+### With Aligned VAD Thresholds
+
+When using AssemblyAI with an external VAD (e.g., Silero VAD), align both thresholds for optimal performance:
+
+```python
+from pipecat.audio.vad.silero import SileroVADAnalyzer, VADParams
+from pipecat.processors.aggregators.llm_response_universal import (
+    LLMContextAggregatorPair,
+    LLMUserAggregatorParams,
+)
+from pipecat.services.assemblyai import AssemblyAISTTService
+from pipecat.services.assemblyai.models import AssemblyAIConnectionParams
+
+# Align VAD thresholds to 0.3 for optimal performance
+stt = AssemblyAISTTService(
+    api_key=os.getenv("ASSEMBLYAI_API_KEY"),
+    connection_params=AssemblyAIConnectionParams(
+        speech_model="u3-rt-pro",
+        vad_threshold=0.3,  # Match with Silero VAD
+    ),
+)
+
+# Configure Silero VAD with matching threshold
+user_aggregator, assistant_aggregator = LLMContextAggregatorPair(
+    context,
+    user_params=LLMUserAggregatorParams(
+        vad_analyzer=SileroVADAnalyzer(
+            params=VADParams(confidence=0.3)  # Match with AssemblyAI
+        )
+    ),
+)
+```
+
 ## Notes
 
-- **u3-rt-pro model**: The default model is now `u3-rt-pro`, which provides the best performance and supports built-in turn detection. 
+- **u3-rt-pro model**: The default model is now `u3-rt-pro`, which provides the best performance and supports built-in turn detection.
+- **VAD threshold alignment**: When using AssemblyAI with external VAD systems (e.g., Silero VAD), align the `vad_threshold` parameter with your VAD's activation threshold (recommended: 0.3 for both) to avoid a "dead zone" where AssemblyAI transcribes speech that your VAD hasn't detected yet. This misalignment can delay interruption handling. AssemblyAI's `vad_threshold` defaults to 0.3, while Pipecat's Silero VAD defaults to 0.7 - we recommend lowering Silero to 0.3 to match.
 - **Turn detection modes**:
   - **Pipecat mode** (`vad_force_turn_endpoint=True`, default): Forces AssemblyAI to return finals ASAP so Pipecat's turn detection (e.g., Smart Turn) decides when the user is done. The service sends a `ForceEndpoint` message when VAD detects the user has stopped speaking.
   - **AssemblyAI mode** (`vad_force_turn_endpoint=False`, u3-rt-pro only): AssemblyAI's model controls turn endings using built-in turn detection. The service emits `UserStartedSpeakingFrame` and `UserStoppedSpeakingFrame` based on AssemblyAI's detection.