Talk Timer

Tracks how long each person talks in a conversation, plus aizuchi (brief listener acknowledgments). No transcription — just timing. Audio never leaves your device.

Speakers

Tap a speaker tile while someone's talking to manually attribute the current bout to them. Bouts under 5 seconds count as aizuchi (acknowledgments) rather than talk time.

Settings

Higher silence threshold = needs louder voice to register. Higher cluster looseness = harder to create separate speakers (good when voices sound similar). Bouts shorter than the min-talk are counted as aizuchi events instead of speech time. Silent gaps shorter than the min-silence don't accrue toward the silence total.

About this

Aizuchi (相槌) is the Japanese term for the short verbal acknowledgments a listener makes while someone else is talking — "mhm," "yeah," "right," "uh-huh." These keep the conversation flowing but aren't really the listener taking the floor. Counting them separately is more honest than either ignoring them or treating them as the listener's "talk time."

How it tells speakers apart: two acoustic features per bout — fundamental frequency (pitch) via autocorrelation, and spectral centroid (brightness) via FFT. Bouts with similar features are grouped into the same speaker. Best with two distinct voices in a quiet room.

Where it fails: very similar voices, heavy crosstalk, background music with speech, or speakerphone audio bleeding back through the mic. Use the manual override (tap a speaker tile while a bout is in progress) when needed.

Privacy: all processing — VAD, pitch, FFT, clustering — runs in the browser. No audio or data leaves the device.