Together AI voice team lead Rishabh Bhargava describes a cascading architecture that streams speech through transcription, a language model and speech synthesis. He discusses turn detectionTurn-taking coordinates when each participant speaks in a conversation., latency and accurate recognition of important words.
Bhargava examines time to first token and audio, network proximity and scaling long-lived conversations. He presents latency figures as examples rather than universal production guarantees.
He contrasts the pipeline with speech-to-speech models, discussing potential gains in conversational nuance alongside instruction-following, tool-calling and evaluation challengesEvaluation is the systematic assessment of a system's behaviour or results against explicit criteria.. Audience questions cover guardrails, model placement and observability.
Watch on YouTube




