It’s not really that Deepgram is broken it’s that phone-call audio + real-time decoding makes it guess a lot, especially for names and numbers. In Bahasa it gets worse because the model relies on context, so it sometimes “fills in” the wrong phrase when it’s unsure. If you want it more stable, you usually fix it by tuning the streaming settings and adding a small cleanup step after transcription rather than just switching STT again.