rasa-pro 3.11.0a4.dev3__py3-none-any.whl → 3.11.0rc2__py3-none-any.whl

rasa/core/channels/voice_stream/browser_audio.py

@@ -1,7 +1,10 @@
 import audioop
+import base64
+import json
+
 import structlog
 import uuid
-from typing import Any, Awaitable, Callable, List, Optional
+from typing import Any, Awaitable, Callable, Optional, Tuple
 
 from sanic import Blueprint, HTTPResponse, Request, response
 from sanic import Websocket  # type: ignore
@@ -9,16 +12,19 @@ from sanic import Websocket  # type: ignore
 
 from rasa.core.channels import UserMessage
 from rasa.core.channels.voice_ready.utils import CallParameters
+from rasa.core.channels.voice_stream.call_state import call_state
 from rasa.core.channels.voice_stream.tts.tts_engine import TTSEngine
 from rasa.core.channels.voice_stream.audio_bytes import RasaAudioBytes
 from rasa.core.channels.voice_stream.voice_channel import (
+    ContinueConversationAction,
+    EndConversationAction,
     NewAudioAction,
     VoiceChannelAction,
     VoiceInputChannel,
     VoiceOutputChannel,
 )
 
-structlogger = structlog.get_logger()
+logger = structlog.get_logger()
 
 
 class BrowserAudioOutputChannel(VoiceOutputChannel):
@@ -31,10 +37,12 @@ class BrowserAudioOutputChannel(VoiceOutputChannel):
     ) -> bytes:
         return audioop.ulaw2lin(rasa_audio_bytes, 4)
 
-    def channel_bytes_to_messages(
-        self, recipient_id: str, channel_bytes: bytes
-    ) -> List[Any]:
-        return [channel_bytes]
+    def channel_bytes_to_message(self, recipient_id: str, channel_bytes: bytes) -> str:
+        return json.dumps({"audio": base64.b64encode(channel_bytes).decode("utf-8")})
+
+    def create_marker_message(self, recipient_id: str) -> Tuple[str, str]:
+        message_id = uuid.uuid4().hex
+        return json.dumps({"marker": message_id}), message_id
 
 
 class BrowserAudioInputChannel(VoiceInputChannel):
@@ -55,8 +63,23 @@ class BrowserAudioInputChannel(VoiceInputChannel):
         self,
         message: Any,
     ) -> VoiceChannelAction:
-        audio_bytes = self.channel_bytes_to_rasa_audio_bytes(message)
-        return NewAudioAction(audio_bytes)
+        data = json.loads(message)
+        if "audio" in data:
+            channel_bytes = base64.b64decode(data["audio"])
+            audio_bytes = self.channel_bytes_to_rasa_audio_bytes(channel_bytes)
+            return NewAudioAction(audio_bytes)
+        elif "marker" in data:
+            if data["marker"] == call_state.latest_bot_audio_id:
+                # Just finished streaming last audio bytes
+                call_state.is_bot_speaking = False  # type: ignore[attr-defined]
+                if call_state.should_hangup:
+                    logger.debug(
+                        "browser_audio.hangup", marker=call_state.latest_bot_audio_id
+                    )
+                    return EndConversationAction()
+            else:
+                call_state.is_bot_speaking = True  # type: ignore[attr-defined]
+        return ContinueConversationAction()
 
     def create_output_channel(
         self, voice_websocket: Websocket, tts_engine: TTSEngine

rasa/core/channels/voice_stream/call_state.py

@@ -0,0 +1,23 @@
+import asyncio
+from contextvars import ContextVar
+from werkzeug.local import LocalProxy
+from dataclasses import dataclass
+from typing import Optional
+
+
+# Per voice session data
+# This is similar to how flask makes the "request" object available as a global variable
+# It's a "global" variable that is local to an async task (i.e. websocket session)
+@dataclass
+class CallState:
+    is_user_speaking: bool = False
+    is_bot_speaking: bool = False
+    silence_timeout_watcher: Optional[asyncio.Task] = None
+    silence_timeout: Optional[float] = None
+    latest_bot_audio_id: Optional[str] = None
+    should_hangup: bool = False
+    connection_failed: bool = False
+
+
+_call_state: ContextVar[CallState] = ContextVar("call_state")
+call_state = LocalProxy(_call_state)

rasa/core/channels/voice_stream/tts/azure.py

@@ -4,7 +4,7 @@ from dataclasses import dataclass
 
 import aiohttp
 import structlog
-from aiohttp import ClientConnectorError
+from aiohttp import ClientConnectorError, ClientTimeout
 
 from rasa.core.channels.voice_stream.audio_bytes import RasaAudioBytes
 from rasa.core.channels.voice_stream.tts.tts_engine import (
@@ -28,10 +28,11 @@ class AzureTTS(TTSEngine[AzureTTSConfig]):
 
     def __init__(self, config: Optional[AzureTTSConfig] = None):
         super().__init__(config)
+        timeout = ClientTimeout(total=self.config.timeout)
         # Have to create this class-shared session lazily at run time otherwise
         # the async event loop doesn't work
         if self.__class__.session is None or self.__class__.session.closed:
-            self.__class__.session = aiohttp.ClientSession()
+            self.__class__.session = aiohttp.ClientSession(timeout=timeout)
 
     async def synthesize(
         self, text: str, config: Optional[AzureTTSConfig] = None
@@ -60,6 +61,8 @@ class AzureTTS(TTSEngine[AzureTTSConfig]):
                     raise TTSError(f"TTS failed: {response.text()}")
         except ClientConnectorError as e:
             raise TTSError(e)
+        except TimeoutError as e:
+            raise TTSError(e)
 
     @staticmethod
     def get_request_headers() -> dict[str, str]:
@@ -92,6 +95,7 @@ class AzureTTS(TTSEngine[AzureTTSConfig]):
         return AzureTTSConfig(
             language="en-US",
             voice="en-US-JennyNeural",
+            timeout=10,
             speech_region="germanywestcentral",
         )
 

rasa/core/channels/voice_stream/tts/cartesia.py

@@ -3,13 +3,13 @@ from typing import AsyncIterator, Dict, Optional
 import os
 import aiohttp
 import structlog
-from aiohttp import ClientConnectorError
+from aiohttp import ClientConnectorError, ClientTimeout
 
 from rasa.core.channels.voice_stream.tts.tts_engine import (
     TTSEngineConfig,
 )
 
-from rasa.core.channels.voice_stream.audio_bytes import RasaAudioBytes
+from rasa.core.channels.voice_stream.audio_bytes import HERTZ, RasaAudioBytes
 from rasa.core.channels.voice_stream.tts.tts_engine import TTSEngine, TTSError
 from rasa.shared.exceptions import ConnectionException
 
@@ -29,10 +29,11 @@ class CartesiaTTS(TTSEngine[CartesiaTTSConfig]):
 
     def __init__(self, config: Optional[CartesiaTTSConfig] = None):
         super().__init__(config)
+        timeout = ClientTimeout(total=self.config.timeout)
         # Have to create this class-shared session lazily at run time otherwise
         # the async event loop doesn't work
         if self.__class__.session is None or self.__class__.session.closed:
-            self.__class__.session = aiohttp.ClientSession()
+            self.__class__.session = aiohttp.ClientSession(timeout=timeout)
 
     @staticmethod
     def get_tts_endpoint() -> str:
@@ -55,13 +56,13 @@ class CartesiaTTS(TTSEngine[CartesiaTTSConfig]):
             "output_format": {
                 "container": "raw",
                 "encoding": "pcm_mulaw",
-                "sample_rate": 8000,
+                "sample_rate": HERTZ,
             },
         }
 
     @staticmethod
     def get_request_headers(config: CartesiaTTSConfig) -> dict[str, str]:
-        cartesia_api_key = os.environ.get(CARTESIA_API_KEY)
+        cartesia_api_key = os.environ[CARTESIA_API_KEY]
         return {
             "Cartesia-Version": str(config.version),
             "Content-Type": "application/json",
@@ -88,13 +89,15 @@ class CartesiaTTS(TTSEngine[CartesiaTTSConfig]):
                     return
                 else:
                     structlogger.error(
-                        "azure.synthesize.rest.failed",
+                        "cartesia.synthesize.rest.failed",
                         status_code=response.status,
                         msg=response.text(),
                     )
                     raise TTSError(f"TTS failed: {response.text()}")
         except ClientConnectorError as e:
             raise TTSError(e)
+        except TimeoutError as e:
+            raise TTSError(e)
 
     def engine_bytes_to_rasa_audio_bytes(self, chunk: bytes) -> RasaAudioBytes:
         """Convert the generated tts audio bytes into rasa audio bytes."""
@@ -105,6 +108,7 @@ class CartesiaTTS(TTSEngine[CartesiaTTSConfig]):
         return CartesiaTTSConfig(
             language="en",
             voice="248be419-c632-4f23-adf1-5324ed7dbf1d",
+            timeout=10,
             model_id="sonic-english",
             version="2024-06-10",
         )

rasa/core/channels/voice_stream/tts/tts_engine.py

@@ -18,6 +18,7 @@ E = TypeVar("E", bound="TTSEngine")
 class TTSEngineConfig(MergeableConfig):
     language: Optional[str] = None
     voice: Optional[str] = None
+    timeout: Optional[int] = None
 
 
 class TTSEngine(Generic[T]):

rasa/core/channels/voice_stream/twilio_media_streams.py

@@ -1,15 +1,17 @@
 import base64
 import json
-import structlog
-from typing import Any, Awaitable, Callable, Dict, List, Optional, Text
 import uuid
 
+import structlog
+from typing import Any, Awaitable, Callable, Dict, Optional, Text, Tuple
+
 from sanic import Blueprint, HTTPResponse, Request, response
 from sanic import Websocket  # type: ignore
 
 
 from rasa.core.channels import UserMessage
 from rasa.core.channels.voice_ready.utils import CallParameters
+from rasa.core.channels.voice_stream.call_state import call_state
 from rasa.core.channels.voice_stream.tts.tts_engine import TTSEngine
 from rasa.core.channels.voice_stream.audio_bytes import RasaAudioBytes
 from rasa.core.channels.voice_stream.voice_channel import (
@@ -21,7 +23,7 @@ from rasa.core.channels.voice_stream.voice_channel import (
     VoiceOutputChannel,
 )
 
-structlogger = structlog.get_logger()
+logger = structlog.get_logger(__name__)
 
 
 def map_call_params(data: Dict[Text, Any]) -> CallParameters:
@@ -47,10 +49,18 @@ class TwilioMediaStreamsOutputChannel(VoiceOutputChannel):
     ) -> bytes:
         return base64.b64encode(rasa_audio_bytes)
 
-    def channel_bytes_to_messages(
-        self, recipient_id: str, channel_bytes: bytes
-    ) -> List[Any]:
+    def create_marker_message(self, recipient_id: str) -> Tuple[str, str]:
         message_id = uuid.uuid4().hex
+        mark_message = json.dumps(
+            {
+                "event": "mark",
+                "streamSid": recipient_id,
+                "mark": {"name": message_id},
+            }
+        )
+        return mark_message, message_id
+
+    def channel_bytes_to_message(self, recipient_id: str, channel_bytes: bytes) -> str:
         media_message = json.dumps(
             {
                 "event": "media",
@@ -60,15 +70,7 @@ class TwilioMediaStreamsOutputChannel(VoiceOutputChannel):
                 },
             }
         )
-        mark_message = json.dumps(
-            {
-                "event": "mark",
-                "streamSid": recipient_id,
-                "mark": {"name": message_id},
-            }
-        )
-        self.latest_message_id = message_id
-        return [media_message, mark_message]
+        return media_message
 
 
 class TwilioMediaStreamsInputChannel(VoiceInputChannel):
@@ -103,9 +105,16 @@ class TwilioMediaStreamsInputChannel(VoiceInputChannel):
         elif data["event"] == "stop":
             return EndConversationAction()
         elif data["event"] == "mark":
-            if data["mark"]["name"] == self.hangup_after:
-                structlogger.debug("twilio_streams.hangup", marker=self.hangup_after)
-                return EndConversationAction()
+            if data["mark"]["name"] == call_state.latest_bot_audio_id:
+                # Just finished streaming last audio bytes
+                call_state.is_bot_speaking = False  # type: ignore[attr-defined]
+                if call_state.should_hangup:
+                    logger.debug(
+                        "twilio_streams.hangup", marker=call_state.latest_bot_audio_id
+                    )
+                    return EndConversationAction()
+            else:
+                call_state.is_bot_speaking = True  # type: ignore[attr-defined]
         return ContinueConversationAction()
 
     def create_output_channel(

rasa/core/channels/voice_stream/util.py

@@ -5,7 +5,7 @@ from typing import Optional, Type, TypeVar
 
 import structlog
 
-from rasa.core.channels.voice_stream.audio_bytes import RasaAudioBytes
+from rasa.core.channels.voice_stream.audio_bytes import HERTZ, RasaAudioBytes
 from rasa.shared.exceptions import RasaException
 
 structlogger = structlog.get_logger()
@@ -23,16 +23,16 @@ def read_wav_to_rasa_audio_bytes(file_name: str) -> Optional[RasaAudioBytes]:
         wave_data = audioop.lin2lin(wave_data, wave_object.getsampwidth(), 1)
         # 8 bit is unsigned
         # wave_data = audioop.bias(wave_data, 1, 128)
-    if wave_object.getframerate() != 8000:
+    if wave_object.getframerate() != HERTZ:
         wave_data, _ = audioop.ratecv(
-            wave_data, 1, 1, wave_object.getframerate(), 8000, None
+            wave_data, 1, 1, wave_object.getframerate(), HERTZ, None
         )
     wave_data = audioop.lin2ulaw(wave_data, 1)
     return RasaAudioBytes(wave_data)
 
 
 def generate_silence(length_in_seconds: float = 1.0) -> RasaAudioBytes:
-    return RasaAudioBytes(b"\00" * int(length_in_seconds * 8000))
+    return RasaAudioBytes(b"\00" * int(length_in_seconds * HERTZ))
 
 
 T = TypeVar("T", bound="MergeableConfig")

rasa/core/channels/voice_stream/voice_channel.py

@@ -1,25 +1,48 @@
 import asyncio
-import logging
+import structlog
 import copy
 from dataclasses import asdict, dataclass
-from typing import Any, Awaitable, Callable, Dict, List, Optional
+from typing import Any, AsyncIterator, Awaitable, Callable, Dict, List, Optional, Tuple
+
+from rasa.core.channels.voice_stream.util import generate_silence
+from rasa.shared.core.constants import (
+    SILENCE_TIMEOUT_DEFAULT_VALUE,
+    SLOT_SILENCE_TIMEOUT,
+)
+from rasa.shared.utils.common import (
+    class_from_module_path,
+    mark_as_beta_feature,
+)
+from rasa.shared.utils.cli import print_error_and_exit
 
 from sanic.exceptions import ServerError, WebsocketClosed
 
 from rasa.core.channels import InputChannel, OutputChannel, UserMessage
 from rasa.core.channels.voice_ready.utils import CallParameters
+from rasa.core.channels.voice_ready.utils import validate_voice_license_scope
 from rasa.core.channels.voice_stream.asr.asr_engine import ASREngine
-from rasa.core.channels.voice_stream.asr.asr_event import ASREvent, NewTranscript
+from rasa.core.channels.voice_stream.asr.asr_event import (
+    ASREvent,
+    NewTranscript,
+    UserStartedSpeaking,
+)
 from sanic import Websocket  # type: ignore
 
 from rasa.core.channels.voice_stream.asr.deepgram import DeepgramASR
-from rasa.core.channels.voice_stream.audio_bytes import RasaAudioBytes
+from rasa.core.channels.voice_stream.asr.azure import AzureASR
+from rasa.core.channels.voice_stream.audio_bytes import HERTZ, RasaAudioBytes
+from rasa.core.channels.voice_stream.call_state import (
+    CallState,
+    _call_state,
+    call_state,
+)
 from rasa.core.channels.voice_stream.tts.azure import AzureTTS
 from rasa.core.channels.voice_stream.tts.tts_engine import TTSEngine, TTSError
 from rasa.core.channels.voice_stream.tts.cartesia import CartesiaTTS
 from rasa.core.channels.voice_stream.tts.tts_cache import TTSCache
+from rasa.utils.io import remove_emojis
 
-logger = logging.getLogger(__name__)
+logger = structlog.get_logger(__name__)
 
 
 @dataclass
@@ -43,25 +66,55 @@ class ContinueConversationAction(VoiceChannelAction):
 
 
 def asr_engine_from_config(asr_config: Dict) -> ASREngine:
-    name = str(asr_config["name"]).lower()
+    name = str(asr_config["name"])
     asr_config = copy.copy(asr_config)
     asr_config.pop("name")
-    if name == "deepgram":
+    if name.lower() == "deepgram":
         return DeepgramASR.from_config_dict(asr_config)
+    if name == "azure":
+        return AzureASR.from_config_dict(asr_config)
     else:
-        raise NotImplementedError
+        mark_as_beta_feature("Custom ASR Engine")
+        try:
+            asr_engine_class = class_from_module_path(name)
+            return asr_engine_class.from_config_dict(asr_config)
+        except NameError:
+            print_error_and_exit(
+                f"Failed to initialize ASR Engine with type '{name}'. "
+                f"Please make sure the method `from_config_dict`is implemented."
+            )
+        except TypeError as e:
+            print_error_and_exit(
+                f"Failed to initialize ASR Engine with type '{name}'. "
+                f"Invalid configuration provided. "
+                f"Error: {e}"
+            )
 
 
 def tts_engine_from_config(tts_config: Dict) -> TTSEngine:
-    name = str(tts_config["name"]).lower()
+    name = str(tts_config["name"])
     tts_config = copy.copy(tts_config)
     tts_config.pop("name")
-    if name == "azure":
+    if name.lower() == "azure":
         return AzureTTS.from_config_dict(tts_config)
-    elif name == "cartesia":
+    elif name.lower() == "cartesia":
         return CartesiaTTS.from_config_dict(tts_config)
     else:
-        raise NotImplementedError(f"TTS engine {name} is not implemented")
+        mark_as_beta_feature("Custom TTS Engine")
+        try:
+            tts_engine_class = class_from_module_path(name)
+            return tts_engine_class.from_config_dict(tts_config)
+        except NameError:
+            print_error_and_exit(
+                f"Failed to initialize TTS Engine with type '{name}'. "
+                f"Please make sure the method `from_config_dict`is implemented."
+            )
+        except TypeError as e:
+            print_error_and_exit(
+                f"Failed to initialize ASR Engine with type '{name}'. "
+                f"Invalid configuration provided. "
+                f"Error: {e}"
+            )
 
 
 class VoiceOutputChannel(OutputChannel):
@@ -71,70 +124,142 @@ class VoiceOutputChannel(OutputChannel):
         tts_engine: TTSEngine,
         tts_cache: TTSCache,
     ):
+        super().__init__()
         self.voice_websocket = voice_websocket
         self.tts_engine = tts_engine
         self.tts_cache = tts_cache
 
-        self.should_hangup = False
         self.latest_message_id: Optional[str] = None
 
     def rasa_audio_bytes_to_channel_bytes(
         self, rasa_audio_bytes: RasaAudioBytes
     ) -> bytes:
+        """Turn rasa's audio byte format into the format for the channel."""
+        raise NotImplementedError
+
+    def channel_bytes_to_message(self, recipient_id: str, channel_bytes: bytes) -> str:
+        """Wrap the bytes for the channel in the proper format."""
         raise NotImplementedError
 
-    def channel_bytes_to_messages(
-        self, recipient_id: str, channel_bytes: bytes
-    ) -> List[Any]:
+    def create_marker_message(self, recipient_id: str) -> Tuple[str, str]:
+        """Create a marker message for a specific channel."""
         raise NotImplementedError
 
+    async def send_marker_message(self, recipient_id: str) -> None:
+        """Send a message that marks positions in the audio stream."""
+        marker_message, mark_id = self.create_marker_message(recipient_id)
+        await self.voice_websocket.send(marker_message)
+        self.latest_message_id = mark_id
+
+    def update_silence_timeout(self) -> None:
+        """Updates the silence timeout for the session."""
+        if self.tracker_state:
+            call_state.silence_timeout = (  # type: ignore[attr-defined]
+                self.tracker_state["slots"][SLOT_SILENCE_TIMEOUT]
+            )
+
+    async def send_text_with_buttons(
+        self,
+        recipient_id: str,
+        text: str,
+        buttons: List[Dict[str, Any]],
+        **kwargs: Any,
+    ) -> None:
+        """Uses the concise button output format for voice channels."""
+        await self.send_text_with_buttons_concise(recipient_id, text, buttons, **kwargs)
+
     async def send_text_message(
         self, recipient_id: str, text: str, **kwargs: Any
     ) -> None:
+        text = remove_emojis(text)
+        self.update_silence_timeout()
         cached_audio_bytes = self.tts_cache.get(text)
-
-        if cached_audio_bytes:
-            await self.send_audio_bytes(recipient_id, cached_audio_bytes)
-            return
         collected_audio_bytes = RasaAudioBytes(b"")
-        # Todo: make kwargs compatible with engine config
-        synth_config = self.tts_engine.config.__class__.from_dict({})
-        try:
-            audio_stream = self.tts_engine.synthesize(text, synth_config)
-        except TTSError:
-            # TODO: add message that works without tts, e.g. loading from disc
-            pass
+        seconds_marker = -1
+        if cached_audio_bytes:
+            audio_stream = self.chunk_audio(cached_audio_bytes)
+        else:
+            # Todo: make kwargs compatible with engine config
+            synth_config = self.tts_engine.config.__class__.from_dict({})
+            try:
+                audio_stream = self.tts_engine.synthesize(text, synth_config)
+            except TTSError:
+                # TODO: add message that works without tts, e.g. loading from disc
+                audio_stream = self.chunk_audio(generate_silence())
+
         async for audio_bytes in audio_stream:
             try:
                 await self.send_audio_bytes(recipient_id, audio_bytes)
+                full_seconds_of_audio = len(collected_audio_bytes) // HERTZ
+                if full_seconds_of_audio > seconds_marker:
+                    await self.send_marker_message(recipient_id)
+                    seconds_marker = full_seconds_of_audio
+
             except (WebsocketClosed, ServerError):
                 # ignore sending error, and keep collecting and caching audio bytes
-                self.should_hangup = True
-
+                call_state.connection_failed = True  # type: ignore[attr-defined]
             collected_audio_bytes = RasaAudioBytes(collected_audio_bytes + audio_bytes)
+        try:
+            await self.send_marker_message(recipient_id)
+        except (WebsocketClosed, ServerError):
+            # ignore sending error
+            pass
+        call_state.latest_bot_audio_id = self.latest_message_id  # type: ignore[attr-defined]
 
-        self.tts_cache.put(text, collected_audio_bytes)
+        if not cached_audio_bytes:
+            self.tts_cache.put(text, collected_audio_bytes)
 
     async def send_audio_bytes(
         self, recipient_id: str, audio_bytes: RasaAudioBytes
     ) -> None:
         channel_bytes = self.rasa_audio_bytes_to_channel_bytes(audio_bytes)
-        for message in self.channel_bytes_to_messages(recipient_id, channel_bytes):
-            await self.voice_websocket.send(message)
+        message = self.channel_bytes_to_message(recipient_id, channel_bytes)
+        await self.voice_websocket.send(message)
+
+    async def chunk_audio(
+        self, audio_bytes: RasaAudioBytes, chunk_size: int = 2048
+    ) -> AsyncIterator[RasaAudioBytes]:
+        """Generate chunks from cached audio bytes."""
+        offset = 0
+        while offset < len(audio_bytes):
+            chunk = audio_bytes[offset : offset + chunk_size]
+            if len(chunk):
+                yield RasaAudioBytes(chunk)
+            offset += chunk_size
+        return
 
     async def hangup(self, recipient_id: str, **kwargs: Any) -> None:
-        self.should_hangup = True
+        call_state.should_hangup = True  # type: ignore[attr-defined]
 
 
 class VoiceInputChannel(InputChannel):
     def __init__(self, server_url: str, asr_config: Dict, tts_config: Dict):
+        validate_voice_license_scope()
         self.server_url = server_url
         self.asr_config = asr_config
         self.tts_config = tts_config
         self.tts_cache = TTSCache(tts_config.get("cache_size", 1000))
 
-        # if set to a value, call will be hungup after marker is reached
-        self.hangup_after: Optional[str] = None
+    async def handle_silence_timeout(
+        self,
+        voice_websocket: Websocket,
+        on_new_message: Callable[[UserMessage], Awaitable[Any]],
+        tts_engine: TTSEngine,
+        call_parameters: CallParameters,
+    ) -> None:
+        timeout = call_state.silence_timeout or SILENCE_TIMEOUT_DEFAULT_VALUE
+        logger.info("voice_channel.silence_timeout_watch_started", timeout=timeout)
+        await asyncio.sleep(timeout)
+        logger.info("voice_channel.silence_timeout_tripped")
+        output_channel = self.create_output_channel(voice_websocket, tts_engine)
+        message = UserMessage(
+            "/silence_timeout",
+            output_channel,
+            call_parameters.stream_id,
+            input_channel=self.name(),
+            metadata=asdict(call_parameters),
+        )
+        await on_new_message(message)
 
     @classmethod
     def from_credentials(cls, credentials: Optional[Dict[str, Any]]) -> InputChannel:
@@ -179,6 +304,7 @@ class VoiceInputChannel(InputChannel):
         channel_websocket: Websocket,
     ) -> None:
         """Pipe input audio to ASR and consume ASR events simultaneously."""
+        _call_state.set(CallState())
         asr_engine = asr_engine_from_config(self.asr_config)
         tts_engine = tts_engine_from_config(self.tts_config)
         await asr_engine.connect()
@@ -192,7 +318,26 @@ class VoiceInputChannel(InputChannel):
 
         async def consume_audio_bytes() -> None:
             async for message in channel_websocket:
+                is_bot_speaking_before = call_state.is_bot_speaking
                 channel_action = self.map_input_message(message)
+                is_bot_speaking_after = call_state.is_bot_speaking
+
+                if not is_bot_speaking_before and is_bot_speaking_after:
+                    logger.info("voice_channel.bot_started_speaking")
+
+                # we just stopped speaking, starting a watcher for silence timeout
+                if is_bot_speaking_before and not is_bot_speaking_after:
+                    logger.info("voice_channel.bot_stopped_speaking")
+                    call_state.silence_timeout_watcher = (  # type: ignore[attr-defined]
+                        asyncio.create_task(
+                            self.handle_silence_timeout(
+                                channel_websocket,
+                                on_new_message,
+                                tts_engine,
+                                call_parameters,
+                            )
+                        )
+                    )
                 if isinstance(channel_action, NewAudioAction):
                     await asr_engine.send_audio_chunks(channel_action.audio_bytes)
                 elif isinstance(channel_action, EndConversationAction):
@@ -232,7 +377,10 @@ class VoiceInputChannel(InputChannel):
     ) -> None:
         """Handle a new event from the ASR system."""
         if isinstance(e, NewTranscript) and e.text:
-            logger.info(f"New transcript: {e.text}")
+            logger.info(
+                "VoiceInputChannel.handle_asr_event.new_transcript", transcript=e.text
+            )
+            call_state.is_user_speaking = False  # type: ignore[attr-defined]
             output_channel = self.create_output_channel(voice_websocket, tts_engine)
             message = UserMessage(
                 e.text,
@@ -242,6 +390,8 @@ class VoiceInputChannel(InputChannel):
                 metadata=asdict(call_parameters),
             )
             await on_new_message(message)
-
-            if output_channel.should_hangup:
-                self.hangup_after = output_channel.latest_message_id
+        elif isinstance(e, UserStartedSpeaking):
+            if call_state.silence_timeout_watcher:
+                call_state.silence_timeout_watcher.cancel()
+                call_state.silence_timeout_watcher = None  # type: ignore[attr-defined]
+            call_state.is_user_speaking = True  # type: ignore[attr-defined]