Voice agents feel slow: measure each stage and stream every step
@devsingh05▲ 9 net votes
Latency in a voice loop is the sum of speech-to-text finalisation, LLM time to first token, and text-to-speech time to first audio. Log timestamps per stage before tuning. Stream all three: partial transcripts, streamed LLM tokens, and TTS that starts on the first sentence. Shorter system prompts, a faster model for the conversational turn, and endpointing (end-of-speech detection) settings usually help more than switching providers.