Deepgram endpointing=300 Cut Off 22% of My Voice AI Users Mid-Answer
The user said, "The biggest trade-off I made on that project was..." and then paused to think. The bot jumped in with "Great, thanks for sharing! Next question." That pause was about 700 milliseconds. My Deepgram endpointing was set to 300. So as far as my pipeline knew, the person was done talking. I had spent two weeks making my voice agent respond faster. It worked. Turns went out quicker than…
Deepgram endpointing sets a silence timer for speech-to-text systems. When no speech is detected for a certain number of milliseconds, Deepgram finalizes the transcript and marks it as speech_final. My voice AI used endpointing=300, which felt fast but cut off 22% of long answers that paused to think. The problem is that 300ms pause time doesn't necessarily indicate the thought is finished.
To fix this, I added a two-gate system: a cheap heuristic to check transcript completeness before committing, along with speculative generation to avoid sluggishness. This reduced cutoffs to 3.1% without sacrificing too much latency. For voice AI, the key is understanding that silence and thought completion are not the same thing.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
