Agent should send interrupt on push_aggregation or EmulateUserStartedSpeakingFrame ?

[seanmuirhead]

Have some cases where the agent is saying the same thing twice in a row without the use getting the chance to speak. I've been able to replicate, and I think it is because we are not pushing StartInterruptionFrames when we push EmulateUserStartedSpeakingFrame and when we are pushing an aggregation from the LLMUserContextAggregator?

[aconchillo]

Have some cases where the agent is saying the same thing twice in a row without the use getting the chance to speak. I've been able to replicate, and I think it is because we are not pushing StartInterruptionFrames when we push EmulateUserStartedSpeakingFrame and when we are pushing an aggregation from the LLMUserContextAggregator?

In theory, if a EmulateUserStartedSpeakingFrame is pushed upstream this will immediately cause an interruption: StartInterruptionFrame + UserStartedSpeakingFrame.

Can you describe Pipecat version and what processors do you have in your pipeline? For example, do you use a STTMuteFilter?

[seanmuirhead]

In theory, if a EmulateUserStartedSpeakingFrame is pushed upstream this will immediately cause an interruption: StartInterruptionFrame + UserStartedSpeakingFrame.

Ah makes sense

Can you describe Pipecat version and what processors do you have in your pipeline? For example, do you use a STTMuteFilter?

Version: 0.0.57

Yes we do use the STTMuteFilter, here is our Pipeline:

pipeline = Pipeline(
        [
            transport.input(),
            stt_mute_filter,  # Should always be before STT
            stt,
            context_aggregator.user(),
            llm,
            tts,
            custom_frame_processor,
            transport.output(),
            context_aggregator.assistant(),
        ],
    )

    task = PipelineTask(pipeline, params=PipelineParams(allow_interruptions=True))

The custom_frame_processor controls the stt_mute_filter sometimes, depending on a variety of TTS events