Why Voice AI Agents Interrupt People: Measured on a Live API and Fixed with Russian Syntax Rules

An engineer measured the delay in voice AI responses using a synthesized Russian phrase through the Yandex Realtime API (compatible with OpenAI Realtime). They varied the silence_duration_ms parameter: with 400 ms or 800 ms, the agent responded almost instantly but only heard the greeting, missing the actual question. With 1200 ms, the full phrase was recognized, but the delay was 1275 ms, of which only 75 ms was model processing. They built a rule-based detector for Russian syntax that decides the appropriate silence threshold: 250 ms if the utterance is likely complete, 1400 ms if incomplete (e.g., ending with prepositions or conjunctions), and 700 ms if uncertain. On a test set of 72 real phone-style phrases, this approach achieved 97% accuracy (70 out of 72), reducing interruptions from 38 to 2 and average silence before answers from 1200 ms to 316 ms. The remaining two failures are pragmatics (e.g., 'no, you misunderstood me'), which rules cannot handle; the author plans to compare with a trained semantic model. They also warn to trim leading/trailing silence when measuring delays on synthesized speech, otherwise results are skewed.