You can record agent and customer audio in separate channels using LiveKit's dual-channel composite recording feature. When starting a room composite egress, set the

audio_mixing

mode to DUAL_CHANNEL_AGENT. This will place the agent's audio in the left channel and all other (customer) audio in the right channel. Here is a Python example:

from livekit import api

req = api.RoomCompositeEgressRequest(
    room_name="your-room-name",
    audio_mixing=AudioMixing.DUAL_CHANNEL_AGENT,  # Agent left, customers right
    file_outputs=[api.EncodedFileOutput(
        file_type=api.EncodedFileType.OGG,
        filepath="recording.ogg",
        # ... your storage configuration
    )]
)

lkapi = api.LiveKitAPI()
res = await lkapi.egress.start_room_composite_egress(req)

This approach creates an OGG file with two channels, enabling tools like Deepgram to provide word-level timestamps per channel. Alternatively, if you want to record agent and customer completely separately (as individual files), you can use Participant Egress: record each participant by specifying their identity separately in the API request. This produces one file per participant. Either approach will let you perform more reliable transcription and speaker separation. Let me know if you need step-by-step guidance for a specific method or integration! Reference: Sources: livekit-protocol/src/livekit.rs | src/main/kotlin/io/livekit/server/EgressServiceClient.kt | Recording Participants | LiveKit Docs | livekit-agents/livekit/agents/voice/room_io/room_io.py | How to set a fixed participant for agent audio processing | Session recording and transcripts | LiveKit Docs