—
Preparing…
First run downloads ~50–100 MB of model weights. Cached by your browser afterwards.
no audio yet
Tip: short clips (5–30 s) transcribe fastest. Recording stops on second click; say a sentence with a clear pause in the middle to see pause markers.
Word timeline
Click a word to seek the audio there · words highlight during playback · ⋯ = pause > 500 ms (gap to previous word)
Why word timing matters for voice agents
- Barge-in detection: when the agent is speaking and the user starts mid-sentence, word-level onset times tell the agent where the user cut in and what to do with the tail of its own utterance (keep, rewind, or drop).
- Subtitle alignment: word timestamps are exactly what caption/overlay systems need to light up each word as it is spoken — no forced alignment post-pass required.
- Latency measurement: comparing the last word's end time against response onset measures end-of-turn → answer latency, the number voice-agent teams obsess over.