I am building a system that listens to a phone conversation and alerts on certain phrases in close to realtime. I’m wondering what is cheaper, should I: A. Use gpt transcribe mini on 2 separate audio channel buffers with silence removed, concat the two transcription results and prompt a chat model every ~10 seconds, keeping a sliding window of like 30s of transcription B. Use gpt realtime mini with transcription, tool calling, and diarization I know realtime is expensive, but I feel like all...