◉ Voice Capability Lab

Standalone test harness for the AI-assessment voice path — in-browser Whisper + the four signal layers. Not the candidate demo.

What this lab answers: can in-browser Whisper deliver per-word timestamps at acceptable latency on this hardware? The ASR Diagnostic panel reports the truth for your machine/backend. If word timestamps don't come back, Layer 2 degrades to a wall-clock estimate and says so — that's a finding, not a failure.

Backend

Model

Ready. Pick a backend & model, then record a ~15s answer.

ASR Diagnostic in-browser Whisper

Backend used—

Model—

Per-word timestamps—

Transcription latency—

Audio decoded—

Layer 1 Transcript (content → Prompt A/B/C, unchanged)

Your transcript will appear here. Filler words are highlighted.

Layer 2 Language

—

proficiency verdict

Speaking rate—

Filler words—

Pauses (>0.5s)—

Vocabulary (TTR)—

strong: WPM 100–170, filler <2/min, TTR >0.55 · adequate: WPM 80–200, filler <4/min, TTR >0.40 · else limited

Layer 3Layer 4 Capture diagnostics

Background noise (SNR)—

Connection type—

Downlink—

Round-trip (RTT)—

Microphone—

Capture sample rate—

Fairness contract: Layers 3–4 are diagnostic only. They set capture_confidence (high / reduced) to inform the recruiter — they never alter the Layer 2 proficiency verdict or any content score. (Spec §13.6 / OD-15.)