cue-v5 / config.json
Joshuant's picture
Cue v5: real conversation (AMI, ICSI) + agent-audio input
117d9d7 verified
Raw History Blame Contribute Delete
1.47 kB
{
"model_type": "cue",
"version": "v5",
"labels": [
"NONE",
"USER_TURN",
"BACKCHANNEL",
"ACCIDENTAL",
"SOFT_INTERRUPT",
"HARD_INTERRUPT",
"TAKEOVER"
],
"encoder": "encoder",
"encoder_type": "whisper",
"encoder_source": "openai/whisper-small (encoder only, positions cut to 100)",
"positions": "none",
"encoder_dim": 768,
"hop_ms": 20,
"chunk_ms": 160,
"context_ms": 2000,
"d": 384,
"layers": 6,
"heads": 4,
"window_frames": 80,
"policy": {
"barge_in": {
"stop_p": 0.8,
"hold": 6,
"cooldown_ms": 1500
},
"end_of_turn": {
"p_done": 0.9,
"turn_hold": 12,
"min_speech": 10,
"fallback_ms": 2000
},
"profiles": {
"responsive": {
"stop_p": 0.7,
"hold": 4,
"p_done": 0.5,
"turn_hold": 6
},
"balanced": {
"stop_p": 0.8,
"hold": 6,
"p_done": 0.9,
"turn_hold": 12
},
"cautious": {
"stop_p": 0.95,
"hold": 6,
"p_done": 0.95,
"turn_hold": 16
}
}
},
"training": {
"init": "Cue v4",
"steps": 8000,
"sources": "['syn=runs/v5/feat/syn:0.5', 'ami=runs/v5/feat/ami:0.3', 'icsi=runs/v5/feat/icsi:0.2']",
"data": "synthetic Indian English calls (clean + realistic; 32 with Dia-voiced callers), AMI and ICSI meetings (caller side only), MUSAN noise"
},
"two_channel": true,
"inputs": "caller audio + the agent's own audio (what it is playing; zero if not given) + agent-speaking flag"
}