the audio files generated by livekit of recorded p...
# ask-aic
calm-hamburger-67669
12/22/2025, 5:18 PMthe audio files generated by livekit of recorded phone calls have a wrong meta tag of duration . Example, an audio of 2 minutes, the meta tag define 4 minutes and it have impact in our UI when we put a player to reproduce the audio. Is this a known error?
t
tall-airline-16756
12/22/2025, 5:18 PMThanks for reaching out, Mauricio. I'm an AI assistant and I'm here to help. Give me a moment while I research this topic.
tall-airline-16756
12/22/2025, 5:19 PMThanks — I’m sorry this is causing trouble in your UI. You’re asking whether this incorrect duration metadata on LiveKit-generated call recordings is a known issue.
Short answer: This is not documented as a known LiveKit bug. LiveKit’s egress produces audio (e.g., Ogg for Opus) and a JSON manifest by default, but there’s no explicit note that embedded duration tags can be wrong.
What can cause this and how to troubleshoot
• Container/codec behaviour: VBR encoding or incomplete frame finalization can make container-level duration tags differ from the real playback length.
• Manifest vs embedded tag: The egress JSON manifest contains metadata; the file’s internal duration tag and the manifest can disagree.
• Player interpretation: Some audio players rely on the embedded tag rather than reading actual samples/frames, producing the UI issue.
• Post-processing: Any processing step after egress (transcoding, concatenation) can alter or leave incorrect duration tags.
Quick steps you can try
1. Compare durations:
◦ Check the egress JSON manifest duration (if present).
◦ Inspect actual media duration with a media tool (ffprobe/mediainfo) and compare to the embedded tag your player uses.
1. Re-mux or normalize the file:
◦ Use ffmpeg to re-multiplex the file without re-encoding:
ffmpeg -i input.ogg -c copy output.ogge
enough-match-43171
12/22/2025, 6:40 PMHow can i confirm that agent is seeing the current shared screen correctly, how would i know ?
t
tall-airline-16756
12/22/2025, 6:40 PMThanks — I’m responding to your question about how to confirm an agent is seeing the current shared screen.
You can’t directly read “what the agent is seeing,” but you can reliably confirm they are receiving and rendering the shared screen by these checks:
1. Verify the agent subscribed to the screen-share track
• Listen for the
track_subscribedpublication.sourceSCREEN_SHAREkindvideocanSubscribe: truetrack_subscribede
enough-match-43171
12/22/2025, 6:41 PMI am using python sdk, give me code snippets to verify these claims
t
tall-airline-16756
12/22/2025, 6:42 PMThanks — responding with Python SDK snippets to:
• detect subscription to a screen-share track,
• confirm the track source/kind,
• count/monitor incoming video frames (confirm rendering/receipt),
• and expose a simple "recent-frame" status you can surface in UI.
1) Listen for track_subscribed and check source/kind
Copy code
# assumes an existing Room connection `room` from livekit SDK
from livekit import TrackSource, RemoteTrackPublication
def on_track_subscribed(track, publication: RemoteTrackPublication, participant):
# publication.source will be a TrackSource (e.g., SCREEN_SHARE)
print("track subscribed:", publication.sid, "source:", publication.source, "kind:", publication.kind)
if publication.source == TrackSource.SCREEN_SHARE and publication.kind == "video":
print(f"Subscribed to screen share from {participant.identity}")
# start monitoring frames for this track
monitor_video_track(track, publication)
room.on("track_subscribed", on_track_subscribed) Copy code
import asyncio
import time
from collections import defaultdict
# store state per publication.sid
frame_state = defaultdict(lambda: {"count": 0, "last_ts": 0.0})
def monitor_video_track(track, publication):
# track is a RemoteVideoTrack (async frame iterator)
sid = publication.sid
async def reader():
async for frame in track: # iterates VideoFrame objects
frame_state[sid]["count"] += 1
frame_state[sid]["last_ts"] = time.time()
# optional: small validation, e.g., non-empty frame
if frame.width == 0 or frame.height == 0:
print("Received empty frame (possible issue)")
# if you render to a UI, your renderer callback would also confirm display
# e.g., render_frame_to_ui(frame)
asyncio.create_task(reader()) Copy code
def screen_share_status(publication_sid, stale_seconds=5.0):
s = frame_state.get(publication_sid)
if not s:
return {"status": "no_subscription", "count": 0, "last_ts": None}
age = time.time() - s["last_ts"]
if age < stale_seconds:
return {"status": "receiving", "count": s["count"], "last_ts": s["last_ts"]}
else:
return {"status": "stale", "count": s["count"], "last_ts": s["last_ts"]}roomJoin: truecanSubscribe: true Copy code
{
"video": {
"room": "myroom",
"roomJoin": true,
"canSubscribe": true
}
}track_subscribedscreen_share_statuse
enough-match-43171
12/22/2025, 6:53 PMis user turn completed data hook available in gemini realtime, or do i need to enable the vad for that ?
t
tall-airline-16756
12/22/2025, 6:53 PMThanks — you’re asking whether the
on_user_turn_completedsilero.VADon_user_turn_completed Copy code
from livekit.agents import AgentSession
from livekit.plugins import silero, turn_detector, google
session = AgentSession(
turn_detection=turn_detector.MultilingualModel(), # or use VAD
vad=silero.VAD.load(), # required if using VAD
llm=google.realtime.RealtimeModel(
model="gemini-2.5-flash-native-audio-preview-09-2025",
realtime_input_config=google.types.RealtimeInputConfig(
automatic_activity_detection=google.types.AutomaticActivityDetection(disabled=True),
),
input_audio_transcription=None,
),
stt=google.stt.STT(), # required if using the turn detection model
)on_user_turn_completed3 Views