It’s not just the endpointing_ms. Since it’s phone audio, the main issue is compression + accent variation .....the model is guessing mid-sentence, which is why even words in the middle get mis-transcribed. QA tools catch it because they analyze the full context after, while real-time STT doesn’t have that advantage. So the fix is usually a mix of slightly higher endpointing + better model choice + a light post-correction layer, not just one setting. Are you currently using Deepgram’s nova-2 model or the default one?