Research notes, product updates and engineering write-ups from the team building Tazo.https://tazolabs.comenhttps://tazolabs.com/blog/ask-what-the-moment-needshttps://tazolabs.com/blog/ask-what-the-moment-needsVoice AI agents decide every few hundred ms: respond, wait, yield or escalate. Why most get it wrong, and what an audio-native decision model changes.Fri, 02 Oct 2026 00:00:00 GMTTazo Labsvoice AI agentsvoice AI decision makingreal-time decisionsturn detectionend-of-turn detectionvoice activity detectionVADbackchannelsturn-takingvoice agent architectureaudio-native modelhttps://tazolabs.com/blog/speculative-tool-calling-voice-aihttps://tazolabs.com/blog/speculative-tool-calling-voice-aiTool calls are where voice agent latency goes. Speculative tool calling starts safe lookups while the caller is still talking, without changing behavior.Fri, 02 Oct 2026 00:00:00 GMTTazo Labsspeculative tool callingvoice AI latencyvoice agent tool callsfunction calling latencyspeculative executionspeculative decodingpreemptive generationeager end of turnvoice agent architecturereal-time voice agentshttps://tazolabs.com/blog/voice-ai-decision-modelshttps://tazolabs.com/blog/voice-ai-decision-modelsVoice AI stacks have great models for STT, LLM and TTS. The layer that decides when to speak is still a silence timer. Why it needs a decision model.Fri, 02 Oct 2026 00:00:00 GMTTazo Labsvoice AI stackvoice AI decision modelvoice agent architectureturn-takingend-of-turn detectionvoice activity detectionVADbarge-inbackchannelsfull-duplex speech modelscascaded voice agentsaudio-native model