conversational voice pipeline — every 20 ms frame
mic → ring
20 ms frames
16 kHz mono
VAD
Silero 4.0
< 1 ms
streaming ASR
Parakeet-CTC 0.6B
~150 ms
LLM
first-token
~100 ms
streaming TTS
Kokoro / EL v2.5
~100 ms
speaker
opus
~20 ms
barge-in: VAD cancels LLM + TTS mid-stream
2026 end-to-end latency leaderboard
200
Moshi
320
GPT-4o realtime
500
LiveKit cascade
1500
batch ASR + chat + TTS
2500+
2022 cascaded
rules of thumb
20 ms frames 320 samples @ 16 kHz; everything downstream keeps the cadence
ring buffer 2 s = 32 000 samples; absorbs jitter without adding floor latency
AEC mandatory without it, TTS output loops back into ASR and self-triggers
warm the TTS prime vocoder before first turn or users feel the 150 ms warm-up