adlib

The AI hype man. Build a beat, grab the mic, and he backs you up. Home: adlib.live.

Run it

npm run dev     # http://localhost:3000  (the mic needs localhost or https)
npm test        # beat data, share links, rhyme detection

No build step. Vercel serves public/ as it is and runs api/ as functions.

How it works

public/
  index.html, css/app.css     the page
  js/song.js                  the beat as data: vibes, pattern, share links, dice
  js/engine.js                the sound: clock, mix bus, drums, 808, keys, lead
  js/ears.js + mic-worklet.js the mic: phrases, syllables, the last 10 seconds of you
  js/words.js                 optional speech-to-text, rhyme detection, keyword map
  js/hype.js                  adlib: when he comes in, what he does, how big
  js/stage.js                 the picture: one shader, circles and lines with rings
  js/phrases.js               what each recorded line says (mirrors scripts/phrases.js)
  audio/hype, audio/boom      his recorded voice (ElevenLabs, pre-generated)
  audio/kit                   CC0 TR-808 drums and DX7 keys (see LICENSES.txt)
  data/rhymes.txt             rhyme tails built from the CMU Pronouncing Dictionary
  classic.html                the first version, untouched, at /classic
api/dg-token.js               mints short-lived Deepgram tokens (off until a key is set)
scripts/                      voice generation and the rhyme-table builder

He listens to rhythm first. The beat clock is ours, so he knows where every bar ends. ears.js finds where your phrases start and stop, and hype.js answers in the gap, on the grid. None of that needs speech-to-text, so it works on iPhone.

He doubles you. The tail of your last phrase is sampled from the mic and thrown back on the beat with a tempo-synced echo.

Words are a bonus layer. When a speech engine is running he also puts your word on screen, answers keywords from the voice bank and calls the rhyme. Engines, best first:

1. Deepgram streaming, if DEEPGRAM_API_KEY is set in Vercel. Works on iPhone. Costs money per minute of mic time for every visitor, so put a spending limit on the key.
2. The browser's own recognizer (desktop Chrome and Android), fed our echo-cancelled mic and run on-device where Chrome supports it.
3. Nothing. Rhythm and doubling still work.

Press the backtick key (or add ?debug) for a live readout while you tune.

Voice bank

scripts/phrases.js is what he says. npm run generate renders it through ElevenLabs into public/audio/ (needs ELEVENLABS_API_KEY in .env.local). If you change a line, change it in public/js/phrases.js too so the screen matches the voice.

Look

Circles and straight lines, yellow, red and blue on black, with rings echoing out. The letter grid and ring spacing in stage.js are measured from the adlib mark. No fake paint.