feat: expressive mode

.github/workflows/deploy-examples.yml

@@ -22,7 +22,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        example: [healthcare, survey, frontdesk, drive-thru, inference, avatar]
+        example: [healthcare, survey, frontdesk, drive-thru, inference, avatar, hotel_receptionist]
     steps:
       - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6
         with:

examples/drive-thru/agent.py

@@ -34,6 +34,7 @@
     function_tool,
     inference,
 )
+from livekit.agents.voice import UserStateChangedEvent, presets
 
 load_dotenv()
 
@@ -483,9 +484,17 @@ async def drive_thru_agent(ctx: JobContext) -> None:
                 ],
             },
         ),
-        llm=inference.LLM("openai/gpt-4.1-mini"),
-        tts=inference.TTS("cartesia/sonic-3", voice="f786b574-daa5-4673-aa0c-cbe3e8534c02"),
+        llm=inference.LLM("google/gemma-4-31b-it"),
+        tts=inference.TTS(
+            "inworld/inworld-tts-2",
+            voice="Sarah",
+            extra_kwargs={"delivery_mode": "CREATIVE", "speaking_rate": 1.1},
+        ),
+        expressive=presets.CUSTOMER_SERVICE,
         max_tool_steps=10,
+        # Flip user_state to "away" after 10s of mutual silence so we can
+        # check whether they're still there (default is 15s).
+        user_away_timeout=10.0,
     )
 
     background_audio = BackgroundAudioPlayer(
@@ -543,6 +552,28 @@ async def push_cart() -> None:
 
     userdata.order.on_change = push_cart
 
+    idle_task: asyncio.Task[None] | None = None
+
+    async def _nudge_while_idle() -> None:
+        # Nudge every 10s until the user speaks again — speaking flips
+        # user_state out of "away", which cancels this task below.
+        while True:
+            logger.info("user idle — checking if they're still there")
+            await session.generate_reply(
+                instructions="The user has been idle, see if they're still there"
+            )
+            await asyncio.sleep(10)
+
+    @session.on("user_state_changed")
+    def _on_user_state_changed(ev: UserStateChangedEvent) -> None:
+        nonlocal idle_task
+        if ev.new_state == "away":
+            if idle_task is None or idle_task.done():
+                idle_task = asyncio.create_task(_nudge_while_idle())
+        elif idle_task is not None:
+            idle_task.cancel()
+            idle_task = None
+
     await session.start(agent=DriveThruAgent(userdata=userdata), room=ctx.room)
     await background_audio.start(room=ctx.room, agent_session=session)
 

examples/drive-thru/database.py

@@ -6,49 +6,48 @@
 from pydantic import BaseModel
 
 COMMON_INSTRUCTIONS = (
-    "You are a quick and friendly McDonald’s drive-thru attendant. \n"
-    "Your job is to guide the customer smoothly through their order, speaking in short, natural voice responses. \n"
-    "This is a voice interaction-assume the customer just pulled up and is speaking to you through a drive-thru speaker. \n"
-    "Respond like you're hearing them, not reading text. \n"
-    "Assume they want food, even if they don’t start with a clear request, and help them get what they’re looking for. \n"
+    # Outcome — what a great interaction looks like.
+    "You are Mac, a quick and friendly McDonald’s drive-thru attendant, and a customer has just "
+    "pulled up to the speaker. A great interaction ends with their complete, correct order in the "
+    "ordering system — every item they asked for, at the right size, with nothing they didn’t ask "
+    "for — reached in as few, as natural exchanges as possible. \n"
     "\n\n"
-    "If an item comes in different sizes, always ask for the size unless the customer already gave one. \n"
-    "If a customer orders a 'large meal', automatically assume both the fries and the drink should be large. \n"
-    "Do not ask again to confirm the size of the drink or fries. This inference is meant to streamline the interaction. \n"
-    "If the customer clearly indicates a different size for the fries or drink, respect their preference. \n"
+    # Voice & personality — keep it short and human.
+    "Your output is synthesized directly to speech, so produce a natural verbatim transcript, not "
+    "polished text. Start responses with real reactions (oh, hmm, ah) and fillers (um, uh, like) "
+    'rather than "Absolutely" or "Certainly", and let mid-sentence fillers (like, you know, I '
+    "mean) fall where they naturally would. Use informal phrasing: yeah, gonna, kinda, gotcha, "
+    "lemme. Keep replies short, upbeat, and snappy, and ask about one thing at a time so you never "
+    "overwhelm the customer. Confirm choices warmly ('Alright, one Big Mac Combo!'), and when "
+    "something’s missing or unavailable, say so with empathy and offer the closest option ('Ah, "
+    "we’re out of Sweet Tea right now — can I get you a Coke instead?'). \n"
     "\n\n"
-    "Never infer or assume any detail the customer has not explicitly stated — especially the drink for a combo meal. \n"
-    "If a required detail is missing, always ask the customer before calling any tool. \n"
+    # How to work — infer intent, acknowledge before acting, stop when you have enough.
+    "Assume the customer wants food even if they don’t open with a clear request, and guide them "
+    "toward it. Treat each transcript as a rough draft of what was said — it may contain "
+    "speech-to-text errors, so don’t mention the transcript or repeat its mistakes. When you can "
+    "reasonably infer intent and it’s safe to, just go with it; when the input is genuinely "
+    "ambiguous or nonsensical, ask the customer to repeat. \n"
+    "Before a tool call that takes a moment, give a brief spoken acknowledgment first ('lemme get "
+    "that added') so there’s no dead air. After each step, ask yourself whether you now have "
+    "everything needed to complete the customer’s request: if you do, act; if a required detail "
+    "is still missing, ask for just that one detail. \n"
     "\n\n"
-    "Be fast-keep responses short and snappy. \n"
-    "Sound human-sprinkle in light vocal pauses like 'Mmh…', 'Let me see…', or 'Alright…' at natural moments-but not too often. \n"
-    "Keep everything upbeat and easy to follow. Never overwhelm the customer, don't ask multiple questions at the same time. \n"
-    "\n\n"
-    "When a customer is confused or asks for something that doesn’t exist, let them know politely and suggest something close. \n"
-    "Always confirm what they picked in a warm, clear way, like: 'Alright, one Big Mac Combo!' \n"
-    "If something’s unavailable, say so with empathy: 'Ah, we're out of Sweet Tea right now-can I get you a Coke instead?' \n"
-    "\n\n"
-    "Whenever a customer asks for, changes, or removes something from their order, you MUST use a tool to make it happen. \n"
-    "Don’t fake it. Don’t pretend something was added - actually **call** the tool and make it real on the ordering system. \n"
-    "\n\n"
-    "Transcripts often contain speech-to-text errors-don’t mention the transcript, don’t repeat its mistakes. \n"
-    "Instead treat each user input as a rough draft of what was said. \n"
-    "If you can guess the user’s intent and it’s safe to do so, infer their meaning and respond naturally. \n"
-    "If the transcript is ambiguous/nonsense and you can’t guess their intent, ask the customer to repeat again. \n"
-    "Stay on-topic; if input is nonsensical in a drive-thru context, ask for concise clarification. \n"
-    "\n\n"
-    "Do not add any item on the user's behalf unless they specifically request it. If the user hasn't asked for an item, NEVER add it. \n"
-    "\n\n"
-    "When a customer changes an item or meal, make sure to remove the previous version before adding the new one. \n"
-    "Otherwise, the order may contain duplicates. \n"
-    "\n\n"
-    "Stricly stick to the defined menu, Do not invent or suggest any new sizes or items. \n"
-    "If the item specified by the user is unclear or not **exactly** on the menu, ask for clarification or say you don't have this specific item \n"
-    "E.g: a hamburger isn't a cheeseburger\n"
-    "Do not ask for size unless the item has more than one size option specified. \n"
-    "If an item does not require a size according to the menu, **NEVER** ask the customer to choose one or mention size at all. \n"
-    "\n\n"
-    "If there is any error from the tool, you should inform the customer and ask them to try again."
+    # Hard constraints — these are invariants, not judgment calls.
+    "Constraints that always hold:\n"
+    "- Stick strictly to the defined menu. Never invent items or sizes. If what the customer wants "
+    "isn’t *exactly* on the menu, say you don’t have it and offer the closest match (a hamburger "
+    "isn’t a cheeseburger). \n"
+    "- Any add, change, or removal must go through a tool call — actually call it, never pretend. "
+    "When a customer swaps an item, remove the old one before adding the new so the order has no "
+    "duplicates. \n"
+    "- Only add items the customer explicitly asked for; never add anything on their behalf. \n"
+    "- Don’t assume unstated details — especially the drink in a combo. If a required detail is "
+    "missing, ask before calling the tool. \n"
+    "- Ask about size only for items that actually have more than one size; if an item has a single "
+    "size, don’t mention size at all. For a 'large meal', make both the fries and drink large "
+    "without re-confirming, unless the customer specifies different sizes. \n"
+    "- If a tool returns an error, tell the customer and ask them to try again. \n"
 )
 
 

examples/frontdesk/agent.py

@@ -1,5 +1,6 @@
 from __future__ import annotations
 
+import asyncio
 import datetime
 import logging
 import os
@@ -38,6 +39,7 @@
     task_completion_judge,
     tool_use_judge,
 )
+from livekit.agents.voice import UserStateChangedEvent, presets
 
 load_dotenv()
 
@@ -63,24 +65,51 @@ def __init__(self, *, timezone: str) -> None:
 
         super().__init__(
             instructions=(
-                f"You are Front-Desk, a helpful and efficient voice assistant. "
-                f"Today is {today}. Your main goal is to schedule an appointment for the user. "
-                "This is a voice conversation — speak naturally, clearly, and concisely. "
-                "When the user says hello or greets you, don’t just respond with a greeting — use it as an opportunity to move things forward. "
-                "For example, follow up with a helpful question like: 'Would you like to book a time?' "
-                "When asked for availability, call list_available_slots and offer a few clear, simple options. "
-                "Say things like 'Monday at 2 PM' — avoid timezones, timestamps, and avoid saying 'AM' or 'PM'. "
-                "Use natural phrases like 'in the morning' or 'in the evening', and don’t mention the year unless it’s different from the current one. "
-                "Offer a few options at a time, pause for a response, then guide the user to confirm. "
-                "If the time is no longer available, let them know gently and offer the next options. "
-                "Always keep the conversation flowing — be proactive, human, and focused on helping the user schedule with ease."
+                # Outcome — what a great interaction looks like.
+                f"You are Front-Desk, a helpful and efficient voice assistant. Today is {today}. "
+                "A great interaction ends with the user booked into an appointment slot that works "
+                "for them, reached through a warm, flowing conversation with as little "
+                "back-and-forth as possible. "
+                # Voice & personality — keep it short and human.
+                "Your output is synthesized directly to speech, so produce a natural verbatim "
+                "transcript, not polished text. Start responses with real reactions (oh, hmm, ah) "
+                'and fillers (um, uh, like) rather than "Absolutely" or "Certainly", with '
+                "mid-sentence fillers (like, you know, I mean) where they’d naturally fall. Mirror "
+                "the user's formality: if they're casual, use informal phrasing (gotcha, alright, "
+                "gonna, kinda, lemme, yeah); if they're more formal, keep your speech cleaner. Vary "
+                "your openers across turns — if you opened the last turn with 'gotcha', pick "
+                "'alright' or 'okay' this turn; don't repeat the same opener back-to-back. "
+                # How to work — be proactive, acknowledge before acting, stop when you can move forward.
+                "Be proactive: when the user greets you, use it to move things forward (e.g. "
+                "'Would you like to book a time?') rather than just greeting back. Before a tool "
+                "call that takes a moment, give a brief spoken acknowledgment so there’s no dead "
+                "air. After each result, check whether you can now move the user toward a booking: "
+                "if so, do it; if you're missing something, ask for just that. "
+                # Speaking about times — constraints that keep it natural over voice.
+                "When talking about availability, call list_available_slots and offer a few clear "
+                "options at a time, then pause for a response and guide the user to confirm. Say "
+                "times like 'Monday at 2' — avoid timezones, timestamps, and the words 'AM'/'PM'; "
+                "use natural phrases like 'in the morning' or 'in the evening', and don’t mention "
+                "the year unless it differs from the current one. When listing several times in the "
+                "same window, group them ('in the evening at 4, 5, or 6') instead of repeating the "
+                "time-of-day qualifier on each slot. If a chosen time is no longer available, let "
+                "them know gently and offer the next options."
             )
         )
 
         self._slots_map: dict[str, AvailableSlot] = {}
 
     async def on_enter(self) -> None:
-        await self.session.say("Hello, I can help you schedule an appointment!")
+        hour = datetime.datetime.now(self.tz).hour
+        time_of_day = "morning" if hour < 12 else "afternoon" if hour < 17 else "evening"
+        await self.session.generate_reply(
+            instructions=(
+                f"Say hello and welcome to the caller — it's currently {time_of_day} their time. "
+                "You're the front desk of an office and you're here to help them schedule a visit. "
+                "Invite them to book an appointment to visit, and ask what time works. "
+                "Keep it warm and brief."
+            )
+        )
 
     @function_tool
     async def schedule_appointment(
@@ -100,7 +129,7 @@ async def schedule_appointment(
         email_result = await beta.workflows.GetEmailTask(chat_ctx=self.chat_ctx)
 
         if ctx.speech_handle.interrupted:
-            return
+            return None
 
         ctx.disallow_interruptions()
 
@@ -262,11 +291,41 @@ async def frontdesk_agent(ctx: JobContext):
     session = AgentSession[Userdata](
         userdata=userdata,
         stt=inference.STT("deepgram/nova-3"),
-        llm=inference.LLM("google/gemini-2.5-flash"),
-        tts=inference.TTS("cartesia/sonic-3", voice="39b376fc-488e-4d0c-8b37-e00b72059fdd"),
+        llm=inference.LLM("google/gemma-4-31b-it"),
+        tts=inference.TTS(
+            "inworld/inworld-tts-2",
+            voice="Nadia",
+            extra_kwargs={"delivery_mode": "CREATIVE", "speaking_rate": 1.1},
+        ),
+        expressive=presets.CUSTOMER_SERVICE,
         max_tool_steps=1,
+        # Flip user_state to "away" after 10s of mutual silence so we can
+        # check whether they're still there (default is 15s).
+        user_away_timeout=10.0,
     )
 
+    idle_task: asyncio.Task[None] | None = None
+
+    async def _nudge_while_idle() -> None:
+        # Nudge every 10s until the user speaks again — speaking flips
+        # user_state out of "away", which cancels this task below.
+        while True:
+            logger.info("user idle — checking if they're still there")
+            await session.generate_reply(
+                instructions="The user has been idle, see if they're still there"
+            )
+            await asyncio.sleep(10)
+
+    @session.on("user_state_changed")
+    def _on_user_state_changed(ev: UserStateChangedEvent) -> None:
+        nonlocal idle_task
+        if ev.new_state == "away":
+            if idle_task is None or idle_task.done():
+                idle_task = asyncio.create_task(_nudge_while_idle())
+        elif idle_task is not None:
+            idle_task.cancel()
+            idle_task = None
+
     await session.start(agent=FrontDeskAgent(timezone=timezone), room=ctx.room)
 
 

examples/healthcare/agent.py

@@ -1,3 +1,4 @@
+import asyncio
 import json
 import logging
 import os
@@ -32,6 +33,7 @@
     WarmTransferTask,
 )
 from livekit.agents.llm import ToolError, function_tool
+from livekit.agents.voice import UserStateChangedEvent, presets
 
 logger = logging.getLogger("HealthcareAgent")
 
@@ -563,7 +565,11 @@ def __init__(self, database=None) -> None:
 
     async def on_enter(self) -> None:
         await self.session.generate_reply(
-            instructions="Greet the user and gather the reason for their call."
+            instructions=(
+                "Warmly welcome the user to the healthcare clinic and ask how you can help "
+                'them today, e.g. "Welcome to the healthcare clinic, how can I help you?" '
+                "Then gather the reason for their call."
+            )
         )
 
     async def task_completed_callback(self, event, task_group):
@@ -751,11 +757,41 @@ async def entrypoint(ctx: JobContext):
     session = AgentSession(
         userdata=userdata,
         stt=inference.STT("deepgram/nova-3", language="multi"),
-        llm=inference.LLM("openai/gpt-4.1-mini"),
-        tts=inference.TTS("inworld/inworld-tts-1"),
+        llm=inference.LLM("google/gemma-4-31b-it"),
+        tts=inference.TTS(
+            "inworld/inworld-tts-2",
+            voice="Luna",
+            extra_kwargs={"delivery_mode": "CREATIVE", "speaking_rate": 1.1},
+        ),
+        expressive=presets.HEALTHCARE,
         preemptive_generation=True,
+        # Flip user_state to "away" after 10s of mutual silence so we can
+        # check whether they're still there (default is 15s).
+        user_away_timeout=10.0,
     )
 
+    idle_task: asyncio.Task[None] | None = None
+
+    async def _nudge_while_idle() -> None:
+        # Nudge every 10s until the user speaks again — speaking flips
+        # user_state out of "away", which cancels this task below.
+        while True:
+            logger.info("user idle — checking if they're still there")
+            await session.generate_reply(
+                instructions="The user has been idle, see if they're still there"
+            )
+            await asyncio.sleep(10)
+
+    @session.on("user_state_changed")
+    def _on_user_state_changed(ev: UserStateChangedEvent) -> None:
+        nonlocal idle_task
+        if ev.new_state == "away":
+            if idle_task is None or idle_task.done():
+                idle_task = asyncio.create_task(_nudge_while_idle())
+        elif idle_task is not None:
+            idle_task.cancel()
+            idle_task = None
+
     await session.start(
         agent=HealthcareAgent(database=db),
         room=ctx.room,