I am using a cascaded model It raises `raise RuntimeError Th LiveKit Community #ask-ai

I am using a cascaded model. It raises `raise Runt...

icy-nightfall-64225

07/21/2025, 4:50 PM

I am using a cascaded model. It raises

raise RuntimeError("This generation handle does not allow interruptions")

on the following code

Copy code

self.session.commit_user_turn(
    transcript_timeout=20
)
if self.session.current_speech:
    self.session._activity._current_speech.allow_interruptions = True
    self.session.interrupt()
await self.session.generate_reply(
    user_input=NOT_GIVEN,  # no user input in cascaded model to work correctly
    instructions="Generate an interruption script because the user is speaking too long. ",
    tool_choice=tool_choice,
    allow_interruptions=False,  # has to be False for cascaded API to work correctly
)

Note, I was trying to force the model to generate some speech, even while the user is still speaking (maybe the user is speaking for too long, so we want to stop them.

tall-airline-16756

07/21/2025, 4:50 PM

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

tall-airline-16756

07/21/2025, 4:51 PM

The error

raise RuntimeError("This generation handle does not allow interruptions")

occurs when you attempt to interrupt a speech generation that was created with

allow_interruptions=False

. In your code, this happens because you set

allow_interruptions=False

when calling

generate_reply

, but then try to interrupt the current speech handle, which is not permitted. Key points to consider: • Speech handles created with

allow_interruptions=False

cannot be interrupted. Attempting to do so will always raise this error. • With cascaded models (such as

RealtimeModel

with server-side turn detection),

allow_interruptions=False

is not supported. The system will log a warning and override this setting, but if you force it, you may encounter unexpected behavior. To generate speech while the user is still speaking or to force an interruption: 1. Set

allow_interruptions=True

(the default) when generating speech if you want to allow interruptions. 2. If you want to stop the user and generate a response, you can manually call

interrupt()

on the session, but only if the current speech handle allows interruptions. 3. If you need to prevent interruptions, ensure you do not attempt to interrupt speech handles created with

allow_interruptions=False

. For cascaded models, rely on the server-side turn detection and avoid setting

allow_interruptions=False

. If you need to discard user audio while non-interruptible speech is playing, consider using the

discard_audio_if_uninterruptible

option in your session configuration. Adjust your code to ensure that interruptions are only attempted on speech handles that allow them, and avoid forcing

allow_interruptions=False

with cascaded models. Sources: livekit-agents/livekit/agents/voice/events.py | basics/interrupts_user.py

icy-nightfall-64225

07/21/2025, 4:56 PM

My problem is that

commit_user_turn

immediately generates a reply, and I actually want to interrupt that reply

tall-airline-16756

07/21/2025, 4:57 PM

Hey there! Let me walk you through some friendly ways to control reply generation in your cascaded model. I totally get how tricky this can be, so I'll break it down into some super manageable steps. 1. Override
on_user_turn_completed
and raise
StopResponse
Think of this like a gentle "pause" button. You can stop the automatic reply by implementing this method and raising

StopResponse

. Here's how it looks:

Copy code

async def on_user_turn_completed(self, turn_ctx, new_message):
    # Add your condition here
    raise StopResponse()

This approach is like having a smooth, clean way to hit the brakes on reply generation. 2. Manual interruption with
session.interrupt()
Sometimes you just need to say "hold up!" - and that's exactly what

await session.interrupt()

does. It'll cancel any current or pending replies in a snap. 3. Manual turn control Want more hands-on control? You can manually manage turn detection, only calling

commit_user_turn

when you want a reply. Or use

clear_user_turn()

to skip generating a reply altogether. Pro tip: I'd recommend using the

on_user_turn_completed

hook with

StopResponse

. It gives you the most flexibility and feels super smooth to implement! Sources: Turn detection and interruptions | LiveKit Docs | Pipeline nodes and hooks | LiveKit Docs | livekit-agents/livekit/agents/llm/realtime.py

2 Views

Open in Slack

Previous Next