Tiny Hinglish Turn Detector

Audio-native HOLD vs END decisions at VAD pause checkpoints. Try incomplete phrases, fillers, corrections, and complete Shiprocket-style requests.

⚠️ Development preview. Data scope: preview-only: 3265 rows from train-00010-of-00083; official test remains sealed. No official-test or collected-Hinglish claim is made.

Loaded partial-shard-warmstart-lr3e4-5ep · 151,812 parameters · 4 s suffix window

What to record

Use natural pacing and leave a short pause at the end of each clip. These are prompts, not included evaluation examples.

Expected	Example prompt	Why
HOLD	`mera order number hai... umm...`	filler before missing detail
END	`mera order cancel kar do`	complete request
HOLD	`haan matlab... kal wala parcel...`	self-repair / continuation
END	`haan, kal wala parcel reschedule kar do`	complete after filler
HOLD	`address change karna hai, flat number...`	slot still missing
END	`address change karke Flat 12B kar do`	slot supplied

Current user turn

END threshold

0 1

Silence at checkpoint (ms)

200 1800

Maximum response timeout (ms)

800 3000

Waiting for audio

Decision probabilities

Runtime diagnostics

Interpretation: HOLD means the agent should keep listening. END means it may respond. The production controller also imposes a maximum timeout so uncertain predictions cannot wait forever.