The Lab · instrument 07/09 · interactive

See your voice

A live spectrogram of your own voice — a close cousin of the representation speech models read — computed and drawn entirely on this page, recording nothing.

What's real here

This is the genuine front end of machine hearing: your browser's audio API runs a Fourier transform over the microphone signal many times a second, and the scrolling image plots energy by frequency (log scale, low notes at the bottom) over time. It is the same move a speech model's encoder makes — one mel-scale squash short of the real input. Privacy, stated plainly: the audio is analyzed inside this tab, drawn, and discarded — nothing is recorded, stored, or transmitted, and the microphone stops the moment you press stop. No microphone? The synth demo feeds a generated signal through the identical pipeline.

Where to push

Hum a steady note and watch the harmonic stack — a ladder of bright bands that bunch together as they climb, because the axis is logarithmic; that ladder is the fingerprint of a pitched sound. Slide your hum upward and the whole ladder climbs rigidly with it. Now say a slow "aaah — eeeh — oooh" without changing pitch: the ladder stays put while the bright regions move — those are formants, the resonances that distinguish vowels, and you are watching the difference between pitch and timbre. Hiss an "ssss" and the bottom goes dark while the top lights up: noise, not harmony — which is exactly how a model tells s from z. Whistle for the purest possible picture: a single thin line.

Why this representation won

Everything a transcription model needs is visible in what you're watching, and almost nothing it doesn't: pitch, timbre, timing, and noise separate cleanly, while the raw waveform's sixteen thousand samples a second hide all of it in wiggles. Turning sound into this picture — then treating listening as seeing — is the single most consequential design decision in audio AI, and you now have a feel for exactly what it preserves.

Failure mode

Forgetting what the picture leaves out. A spectrogram this coarse discards phase and fine timing, and a model reading it inherits those blind spots — as well as everything the microphone itself did to your voice: the laptop mic's thin bass, the room's echo, the compressor's pumping. When a speech system misbehaves, the culprit is often upstream of the model, in exactly the stage you've been playing with. The representation is the first place to look, not the last.