redesign interruption strategies

[aconchillo]

Please describe the changes in your PR. If it is addressing an issue, please reference that as well.

Introducing new interruption and speaking strategies. Interruption strategies are used to indicate when the user should interrupt the bot (e.g. using VAD events or when a user says one or more words). Speaking strategies are used to indicate when the bot should start talking (e.g. using an end-of-turn detection model, or checking how long until the last user transcription). In conversational agents these are usually referred to start/stop speaking policies or plans.

Available interruption strategies:

• VADInterruptionStrategy (default)
• MinWordsInterruptionStrategy

Avaliable speaking strategies:

• TranscriptionSpeakingStrategy (default)
• TurnAnalyzerSpeakingStrategy

To set these strategies you do it when setting up the PipelineTask, for example:

task = PipelineTask(..., params=PipelineParams(
    interruption_strategies=[MinWordsInterruptionStrategy(min_words=1)],
    speaking_strategies=[
        TurnAnalyzerSpeakingStrategy(
            turn_analyzer=LocalSmartTurnAnalyzerV3(params=SmartTurnParams())
        )
    ],
))

In order to use the new interruption and speaking strategies you should update to the new universal LLMContext and LLMContextAggregatorPair.

[aconchillo]

It occurred to me that checking for end of sentence also allowed us to use semantic to avoid false positives. This has actually worked really well. Before we were sending requests to the LLM all the time.