I've dealt with this exact issue before - that carrier pattern where they announce the digits is a pain because it bypasses traditional voicemail detection entirely.
There's actually a workaround that combines audio analysis with a custom interrupt handler that catches this specific pattern before your LLM even processes it. Works without adding latency to real conversations.
Happy to walk you through how it works if you want - it's more of a pre-processing layer than a detection model thing.