add support for krisp vad and voice isolation

src/eva/assistant/pipecat_server.py

@@ -568,11 +568,25 @@ async def _on_turn_end(_service, transcript: str) -> None:
 
     def _create_transport(self, websocket) -> FastAPIWebsocketTransport:
         """Create the WebSocket transport with Twilio frame serialization."""
+        audio_in_filter_cfg = self.pipeline_config.audio_in_filter.lower()
+        if audio_in_filter_cfg == "none":
+            audio_in_filter = None
+        elif audio_in_filter_cfg == "krisp_viva":
+            from pipecat.audio.filters.krisp_viva_filter import KrispVivaFilter
+
+            audio_in_filter = KrispVivaFilter(**self.pipeline_config.audio_in_filter_params)
+            logger.info("Using Krisp VIVA audio-in filter")
+        else:
+            raise ValueError(
+                f"Unsupported audio_in_filter: {audio_in_filter_cfg}. Supported types: 'krisp_viva', 'none'"
+            )
+
         return FastAPIWebsocketTransport(
             websocket=websocket,
             params=FastAPIWebsocketParams(
                 audio_in_enabled=True,
                 audio_out_enabled=True,
+                audio_in_filter=audio_in_filter,
                 add_wav_header=False,
                 serializer=TwilioFrameSerializer(
                     self.conversation_id,

src/eva/assistant/pipeline/turn_config.py

@@ -49,8 +49,13 @@ def create_vad_analyzer(vad_type: str, vad_params: dict[str, Any]) -> VADAnalyze
         # Create VADParams, respecting existing defaults if no params specified
         params = VADParams(**vad_params) if vad_params else None
         return SileroVADAnalyzer(params=params)
+    elif vad_type_lower == "krisp_viva":
+        from pipecat.audio.vad.krisp_viva_vad import KrispVivaVadAnalyzer
+
+        params = VADParams(**vad_params) if vad_params else None
+        return KrispVivaVadAnalyzer(params=params)
     else:
-        raise ValueError(f"Unsupported VAD type: {vad_type}. Supported types: 'silero', 'none'")
+        raise ValueError(f"Unsupported VAD type: {vad_type}. Supported types: 'silero', 'none', 'krisp_viva'")
 
 
 def create_turn_start_strategy(

src/eva/models/config.py

@@ -188,7 +188,7 @@ class ModelConfig(BaseModel):
     vad: str = Field(
         "silero",
         description=(
-            "VAD analyzer type: 'silero' or 'none'. Defaults to 'silero' (SileroVADAnalyzer). Use 'none' with external turn strategies (e.g. deepgram-flux) to skip local VAD. Set via EVA_MODEL__VAD."
+            "VAD analyzer type: 'silero', 'krisp_viva' or 'none'. Defaults to 'silero' (SileroVADAnalyzer). Use 'none' with external turn strategies (e.g. deepgram-flux) to skip local VAD. Set via EVA_MODEL__VAD."
         ),
     )
     vad_params: dict[str, Any] = Field(
@@ -198,6 +198,19 @@ class ModelConfig(BaseModel):
         ),
     )
 
+    # Transport-level audio input filter
+    audio_in_filter: str = Field(
+        "none",
+        description=(
+            "Audio input filter applied to the transport: 'krisp_viva' or 'none'. Defaults to "
+            "'none'. Set via EVA_MODEL__AUDIO_IN_FILTER."
+        ),
+    )
+    audio_in_filter_params: dict[str, Any] = Field(
+        {},
+        description="Audio input filter parameters (JSON). Set via EVA_MODEL__AUDIO_IN_FILTER_PARAMS.",
+    )
+
     # CASCADE-only latency controls.
     pre_tool_speech: str = Field(
         "off",

tests/unit/assistant/test_pipecat_server.py

@@ -33,7 +33,6 @@ def _make_server(tmp_path: Path):
     srv._server_task = None
     srv._runner = None
     srv._metrics_observer = None
-    srv._latency_measurements = []
     srv.num_seconds = 0
     srv.pipeline_config = MagicMock()
     srv.conversation_id = "test-conv"

vendor/krisp/README.md

@@ -22,7 +22,9 @@ Result (wheel version and `.kef` set will vary by download):
 vendor/krisp/
 ├── README.md                                          (tracked)
 ├── krisp_audio-1.10.0-cp311-cp311-linux_x86_64.whl    (git-ignored)
-└── krisp-viva-tp-v3.kef                               (git-ignored)
+├── krisp-viva-tp-v3.kef                               (git-ignored)
+├── krisp-viva-vad-v2.kef                              (git-ignored)
+└── krisp-viva-vi-tel-v2.1.kef                         (git-ignored)
 ```
 
 > [!WARNING]
@@ -35,6 +37,8 @@ In the Docker container:
 - The `eva` command (which is actually `scripts/docker_eva_wrapper.sh`) installs the `.whl`, if present.
 - The `KRISP_VIVA_TURN_MODEL_PATH` env var points at the `.kef` file directly on that mount, e.g. `vendor/krisp/krisp-viva-tp-v3.kef`.
 - The `KRISP_VIVA_API_KEY` env var provides the API key.
+- If using the Krisp VAD, set `EVA_MODEL__VAD` to `krisp_viva` and point `KRISP_VIVA_VAD_MODEL_PATH`  to the `krisp-viva-vad-v2.kef` file
+- If using the Krisp audio input filter, set `EVA_MODEL__AUDIO_IN_FILTER` to `krisp_viva` and point `KRISP_VIVA_AUDIO_IN_FILTER_MODEL_PATH` to the `krisp-viva-vi-tel-v2.1.kef` file
 
 ## Runtime requirement