how do I get agent voice transcription for gemini realtime model?

llm=google.realtime.RealtimeModel(
    model="gemini-2.5-flash-native-audio-preview-12-2025",
    voice="Aoede",
    temperature=0.8,

Thanks for reaching out, Gaurav. I'm an AI assistant and I'm here to help. Give me a moment while I research this topic.

@dataclass class TextOutputOptions: sync_transcription: NotGivenOr[bool] = NOT_GIVEN """False to disable transcription synchronization with audio output. Otherwise, transcription is emitted as quickly as available.""" transcription_speed_factor: float = 1.0 """Speed factor of transcription synchronization with audio output. Only effective if

sync_transcription

is True.""" explain with example

You're asking how `TextOutputOptions`'s

sync_transcription

and

transcription_speed_factor

work. Here’s a concise explanation and examples. What each field does •

sync_transcription

(bool): ◦ True (default behavior): transcriptions are timed to match audio playback so captions appear in sync with the spoken words. ◦ False: transcriptions are emitted as soon as they are available (earlier than synced captions). •

transcription_speed_factor

(float): ◦ Only used when

sync_transcription

is True. ◦ Multiplies the rate at which transcription is synchronized to audio. ◦ Values >1.0 make transcription catch up faster (appear slightly earlier relative to audio), values <1.0 delay transcription relative to audio. Examples 1) Synchronous, default speed (captions follow audio normally)

from livekit.plugins import room_io

room_options = room_io.RoomOptions(
    text_output=room_io.TextOutputOptions(
        sync_transcription=True,              # sync with audio
        transcription_speed_factor=1.0        # normal sync speed
    )
)

Behavior: Captions appear aligned with speech timing. 2) Synchronous, faster transcription (captions appear a bit earlier while still aligned)

room_options = room_io.RoomOptions(
    text_output=room_io.TextOutputOptions(
        sync_transcription=True,
        transcription_speed_factor=1.3        # speed up caption delivery
    )
)

Behavior: Captions are still synchronized but delivered slightly sooner to reduce perceived lag. 3) As-fast-as-possible (no sync) — get transcripts immediately

room_options = room_io.RoomOptions(
    text_output=room_io.TextOutputOptions(
        sync_transcription=False               # emit transcripts ASAP
        # transcription_speed_factor ignored
    )
)

Behavior: Transcripts arrive as soon as available (may precede or not align with audio playback). When to use each • Use

sync_transcription=True

for captioning/subtitles where alignment with audio is important. Adjust

transcription_speed_factor

to tune perceived lag. • Use

sync_transcription=False

for real-time text processing, analytics, or when you prefer earliest possible transcripts over precise lip-sync. Want a recommendation for your app (captions vs. realtime analysis)? Tell me your use case and client platform.

2025-12-17 010313,798 - __mp_main__ - INFO - 🤖 Agent state: speaking → listening 010313.798 INFO __mp_main__ 🤖 Agent state: speaking → listening {"pid": 5962, "job_id": "AJ_94JeoHHsyhgj", "room_id": "RM_X8MjttBSZm8q"} 2025-12-17 010313,798 - __mp_main__ - INFO - 💬 assistant: Providing a Response I've determined this message is a standard greeting. It's unrelated to OneAssist services, so the

oneassist_chat

tool isn't needed. I'll craft a friendly, polite response with an Indian accent to address the user appropriately. I am good, thank you! How can I assist you today? INFO __mp_main__ 💬 assistant: Providing a Response I've determined this message is a standard greeting. It's unrelated to OneAssist services, so the

oneassist_chat

tool isn't needed. I'll craft a friendly, polite response with an Indian accent to address the user appropriately. I am good, thank you! How can I assist you today? {"pid": 5962, "job_id": "AJ_94JeoHHsyhgj", "room_id": "RM_X8MjttBSZm8q"} 2025-12-17 010313,798 - __mp_main__ - DEBUG - Skipping thought message: Providing a Response I've determined this mes... DEBUG __mp_main__ Skipping thought message: Providing a Response I've determined this mes... {"pid": 5962, "job_id": "AJ_94JeoHHsyhgj", "room_id": "RM_X8MjttBSZm8q"} this is what I have printed in my logs. is there any way i can get exact transcription of the agent output? for example, i am using session handler as well as room handler

You're already close — to capture the exact text the agent actually spoke (not internal "thought"/debug messages), use one of these, in order of recommendation: 1) conversation_item_added event (recommended — final, clean text)

from livekit.agents import ConversationItemAddedEvent

@session.on("conversation_item_added")
def on_conversation_item(event: ConversationItemAddedEvent):
    if event.item.role == "assistant":
        print("Exact agent transcription:", event.item.text_content)

2) Listen for

lk.transcription

text stream (real-time, from room)

# agent-side example handler for a Room text reader
async def on_text_stream(reader, participant_identity):
    message = await reader.read_all()
    is_final = reader.info.attributes.get("lk.transcription_final") == "true"
    is_agent = reader.info.attributes.get("lk.transcribed_track_id") is not None
    if is_final and is_agent:
        print("Final agent transcription:", message)

3) Override

transcription_node

(capture raw output passed to TTS)

async def transcription_node(self, text: AsyncIterable[str], model_settings):
    async for chunk in text:
        print("Transcription chunk:", chunk)
        yield chunk

4) Session report after the session (batch final transcripts)

report = ctx.make_session_report()
for item in report.conversation:
    if item["role"] == "assistant":
        print("Agent said:", item["text"])

Why your logs show thought/debug lines • Those log entries are LLM/internal reasoning or debug prints, not the committed assistant response. Only

conversation_item_added

,

lk.transcription

final entries,

transcription_node

output, or session reports contain the exact spoken text. If you want, tell me: • Do you want real-time capture (use method 2 or 3) or post-session transcripts (method 1 or 4)? • Are you using the session handler, room handler, or both in your current code so I can give an inline example? Sources: Pipeline nodes and hooks | LiveKit docs | Observability data hooks | LiveKit docs | Events and error handling | LiveKit docs