Nokify

Turn any YouTube song into a monophonic Nokia ringtone — a real RTTTL string you can paste into a ringtone editor, a dumbphone, or an Arduino buzzer.

Paste a link, scrub to the hook, hit NOKIFY. The melody is transcribed in your browser and comes back as the one-line format Nokia shipped in 1998:

Take On Me:d=16,o=5,b=120:f#,f#,d,8b4,8p,8b4,8p,8e,8p,8e,8p,8e,8g#,8g#,8a,8b

No API keys, no model weights, no upload of your own files. The pitch detection runs entirely client-side in Web Audio.

---

How it works

YouTube URL
     │
     ▼  server: python -m yt_dlp        (metadata + cached audio, ≤15 min videos)
     │
     ▼  you scrub a start point, pick 2–20s
     │
     ▼  server: ffmpeg -ss/-t → mono 22.05 kHz WAV, streamed (never written to disk)
     │
     ▼  browser: decodeAudioData → OfflineAudioContext
     │
     ▼  YIN pitch tracking, 100–1100 Hz, threshold 0.15
     │
     ▼  quantize to a 16th-note grid (125 ms cells @ 120 BPM)
     │    majority vote per cell, cell is a rest if <40% of frames are voiced
     │
     ▼  merge equal-pitch runs into notes, drop 1-cell blips between identical pitches
     │
     ▼  RTTTL string  +  A/B playback (square-wave tone vs. the original clip)

The server is 126 lines and does exactly two things: fetch and cut. Everything musical happens in public/app.js.

Why YIN

Autocorrelation alone octave-errors badly on real mixes. YIN's cumulative mean normalized difference function suppresses the sub-harmonic dip that makes a melody jump down an octave mid-phrase. It is still monophonic — it tracks the loudest periodic component, so it locks onto a lead vocal or lead synth and ignores the pad underneath. That limitation is the point: a Nokia ringtone is monophonic.

Why a 16th grid at 120 BPM

RTTTL only encodes durations as fractions of a beat. Snapping to fixed 125 ms cells means every detected note lands on a legal RTTTL duration without a tempo-estimation step that would be wrong half the time on a 10-second excerpt. Songs far from 120 BPM still transcribe — the note lengths quantize, the pitches don't care.

---

Running it

Requirements

• Node 18+
• Python with yt-dlp — pip install yt-dlp (called as python -m yt_dlp)
• ffmpeg on your PATH

npm install
npm start

Open http://localhost:3990.

Downloaded audio is cached in cache/ keyed by video ID, so re-scrubbing the same song doesn't re-download it.

---

The interface

It's a Nokia. Dot-matrix LCD, signal bars, battery meter, scrolling title ticker, two soft keys. This is not a theme layer over a normal form — the LCD is a canvas that draws the detected melody as a piano-roll and sweeps a playhead across it during playback.

Play tone renders the notes through an oscillator. Play original plays the same clip back. Toggling between them is the honest test of whether the transcription is any good.

---

Limits

• Monophonic only. Chords, harmonies and dense mixes transcribe as whichever voice is loudest, which is often not the tune you wanted.
• Best on clear leads. Solo vocal, whistle, lead synth, brass. Worst on distorted guitar and anything with heavy reverb tails.
• Videos longer than 15 minutes are rejected; clips are capped at 20 seconds.
• RTTTL has no concept of velocity, overlap or tempo change. A transcription is a caricature — that's the aesthetic.

---

Repo layout

server.js        Express (3990): /api/info, /api/clip. yt-dlp + ffmpeg.
public/
  index.html     the phone
  app.js         YIN, quantizer, RTTTL encoder, playback, LCD renderer
  style.css
cache/           downloaded audio, gitignored

That is the whole project. The server fetches and cuts; everything musical happens in the browser.

---

License

MIT. See LICENSE.

yt-dlp and ffmpeg are separate projects with their own licenses; this repo does not bundle them. You are responsible for whether your use of a given video is permitted.