lmfvoice
native speech-to-speech knowledge agent · LiquidAI LFM2.5-Audio-1.5B · one model, no cascade · 100% local CPU
connecting…

① Knowledge

empty → BARE turns (model answers from its own weights)

② Say / Ask

type or hold the mic — the model hears and answers in voice
simplex turn-based v1: record → stop → answer. No barge-in yet.

Turn tag

GROUNDED 0
BARE 0
no skip-gate here: LFM is ALWAYS the loop. The tag just says whether knowledge was injected (GROUNDED) or not (BARE).

Live latency

TTFA = time to FIRST audio chunk · RTF = wall-time / audio-seconds (<1.0 beats realtime). /metrics

Transcript

♪ marks audio chunks streaming