I'm building an AI agent in Amazon Connect using the native AI agent option, invoked through the Get customer input block with Lex. The agent is also using Nova Sonic for speech interaction. The problem is that brief acknowledgments such as "okay," "uh-huh," "right," or even small murmurs are treated as full interruptions. Nova Sonic stops speaking, and the agent appears to interpret the sound as a new user prompt, losing its place in the conversation.
I tried disabling interruptions completely, but I still want users to be able to interrupt when they have a meaningful request. Is there an API, configuration, or workaround that can filter out these backchannel responses without turning interruptions off entirely?
2 Answers
If the native block doesn’t expose transcript filtering or interruption thresholds, you may need to move the turn-taking logic outside the fully managed agent flow. For example, place a controlled input or queue layer in front of the agent, classify short acknowledgments, and only send substantive utterances to the AI agent. A queue can also let the agent finish speaking unless the incoming utterance meets a threshold for a real request. With the current black-box setup, disabling interruptions may be the only directly available setting, so the exact workaround depends on which Amazon Connect and Lex extension points your account supports.
A practical approach would be to filter the transcript before it reaches the agent. A lightweight classifier or simple rules could recognize backchannel phrases such as “okay,” “uh-huh,” and “right,” then discard them or treat them as acknowledgments instead of forwarding them as new prompts. Speech-level interruption settings generally control when audio interrupts playback, not whether the recognized text is semantically meaningful. Also check whether the system is using end-of-utterance detection or an aggressive timeout, since that can make short murmurs trigger more often.
That makes sense for a custom voice pipeline, but I’m using the native AI agent inside the Get customer input block. The agent and Nova Sonic behavior seem to be managed internally, and the settings page doesn’t expose much control. I’m not sure whether there’s an API or hook where I could insert that filtering step.
The interaction is handled through the Get customer input block, which calls the AI agent through Lex. I’m trying to find out whether that managed path provides any API or event-level control before I replace it with a custom flow.