CHIPCRUSH

How the sound is made

Chip synth

The song is cut into 17 ms frames and turned into a spectrum. Each possible note gets a score from its overtones (harmonic sum). The best-scoring high note becomes the melody and the best low note becomes the bass. The three strongest pitch classes make the chord, which is played as a fast arpeggio, the old chip way to fake a chord. Voices update 60 times a second with 16 volume steps, like an NES.

Drums and MIDI

A short spectrum watches three bands for sudden energy jumps: lows for kicks, mids for snares and highs for hats. Kicks are falling triangle tones and the rest is LFSR noise. The same notes can be saved as a MIDI file, so you can load them into a tracker or DAW.

Four sound chips

Every chip plays from the same 60 Hz stream of notes. NES: pulse waves, a stepped triangle and LFSR noise. C64 SID: a sawtooth lead with a ring-modulated triangle edge and a pulse arpeggio whose width sweeps, with lead and bass through one resonant state-variable filter that closes after each note. Mega Drive FM: two-operator phase modulation whose depth decays after each note (1:1 slap bass, 1:2 lead with feedback, 1:3.5 chimes) on a 9-bit DAC. PC speaker: one bit; the voices take turns 120 times a second by weighted round robin.

Scale quantizer and note views

The key is found by correlating how long each pitch class sounds with the Krumhansl–Kessler major and minor profiles. Each note then snaps to the nearest note of the chosen scale on that tonic, ties going down, so a major third becomes a minor one. Pattern view shows the result like a FastTracker II pattern: 0.12 s rows, notes, instruments, hex volumes, and chords as 0xy arpeggio effects.

Track decks and loops

Each track plays from its start flag to its loop out point, repeats the loop as many times as set, then plays on, with 3 ms crossfades at the seams. The transport loop instead loops part of the finished mix gaplessly in the player, and can be exported on its own.

Diegetic FX

Battery dying slows the clock along a squared curve so pitch droops, holds samples longer, drops bit depth, adds brown-out gaps and shrinks the headroom. Bad cell keeps 300 Hz to 3.4 kHz, keeps only the strongest partials of each 23 ms frame (the metallic codec sound) and drops 20 ms packets in bursts from a two-state channel, covering them by repeating the last one. Through the wall is a steep low-pass with room modes at 62 and 125 Hz plus a clipped, band-passed buzz driven by the bass.

Story FX

Hot-car cassette reads the tape through one shared, wandering read head: three slow wow sines, a speed sag that grows to 3% as the tape stretches, and random pinches that dip the speed and drop the level and highs. Tunnel radio fades the station on a ramp, flickers it in 40–90 ms steps (multipath), closes a low-pass, and brings in AM static and engine rumble. The diner runs the music through a small Schroeder reverb (four combs, two all-passes) over talkers made of vowel formants. The PA chime is two inharmonic bell tones a major third apart; the voice goes through a 350 Hz–3.8 kHz horn with a 180 ms slap-back while the music ducks.

More sound chips

SNES: one-cycle samples squeezed to 4-bit BRR blocks (16 samples share one power-of-two shift), read through a soft four-point B-spline (the muffled SNES tone), at 32 kHz with a 14-bit pitch register, ADSR envelopes and a 240 ms echo fed back through an eight-tap FIR low-pass. Amiga: 8-bit samples played at whole-number periods of the 3.55 MHz PAL clock with no interpolation, loops shortened for high notes to stay under the 28 kHz DMA limit, 0–64 volume, then the 4.4 kHz RC filter. FM patches set each operator pair's ratio, index sweep, feedback and envelope.

Melody focus

While the melody sounds (with a 0.2 s tail), chords and bass are turned down and, on FM, their modulation index is lowered so they darken. Chord tones within a semitone of the melody's pitch class are dropped, and at high settings a chord that reaches into the melody's register moves down an octave.

Speak & Spell and the chip vocoder

Speak & Spell is LPC as in the TMS5100: 25 ms frames at 8 kHz, ten reflection coefficients from Levinson–Durbin quantized in the arcsine domain to 5, 5, 4, 4, 4, 4, 4, 3, 3 and 3 bits, a 16-step energy and an autocorrelation pitch; a ten-stage lattice filter re-speaks it from a ringing pulse or noise, with parameters stepped eight times a frame. The vocoder splits the recording into 16 bands from 120 Hz to 7 kHz and uses each band's envelope to open the same band of a whitened chip carrier playing the melody, with noise in the top bands.

Tempo, key and DJ mode

Spectral flux at 11 kHz gives an onset curve. Its autocorrelation, weighted toward 120 BPM, gives a rough tempo, refined to 0.02 BPM by sliding a beat grid along the whole song; the beat position with the most low-band onset is the downbeat. The key is Krumhansl–Kessler on the same spectra. DJ mode stretches every deck (WSOLA) to the tuned deck's tempo, counting half- or double-time songs in 8ths or half notes, shifts it the fewest semitones onto the tuned deck's key (minor keys by their relative major), evens out loudness, then rebuilds the mix beat by beat from seeded choices, with 4 ms crossfades at every cut.

Autopilot and ambience

Autopilot decodes one song ahead, renders it with a sound drawn from your taste weights (or one you loved), and sets each song's gain from its gated RMS to about −18 dBFS without letting peaks clip. Ambience beds are generated for the mix's length: filtered stereo noise with slow swells, plus decaying sine ticks for drops, cups and keys, formant babble for the café, and Doppler-shifted engine tones for passing cars.

Bit depth and sample rate

Each sample snaps to one of 2ⁿ levels with round(x · 2ⁿ⁻¹) / 2ⁿ⁻¹ and is held until the chip's slower clock ticks again. Fewer levels add grain. A slower clock drops the highs and folds the rest back as aliasing.

Tempo without pitch

Tempo uses WSOLA. The track is cut into overlapping 42 ms grains, each grain slides up to 12 ms to line up with the one before it, and then they are blended back together. Pitch shift stretches first and then resamples.

Tape and vinyl

Wow and flutter drive a cubic-interpolated fractional delay line with slow capstan sway and fast motor quiver. Dropouts model creased oxide lifting off the tape head, rolling off high frequencies before ducking the volume. Vinyl dust combines a 33⅓ RPM groove swoosh, platter rumble, and impulses through a 2.1 kHz phono pickup resonance.

Robot voice

Ring mod multiplies the sound by a sine tone, which smears every note into metallic sidebands. Metal is a feedback comb filter tuned to the same robot pitch. Glitch flips the chip's lowest bits.

Radio dial

Each station sits on its own spot of a 540 to 1600 kHz AM dial, and all of them play on one timeline from their start flags, looping if they are shorter. As the needle moves off a station, that station fades on a bell curve and gets muffled, a heterodyne whistle rises in pitch, and AM static fills the gaps. Auto tune holds each station, then sweeps on with a little overshoot, like a hand on the knob. The static styles fill the gaps differently: Tuning sweep scatters weak carriers that chirp as their beat tone falls to zero and rises again, plus thin, far-off copies of your other tracks; Shortwave adds Morse, slow fading and sideband voices shifted by a ring modulator; Thunderstorm adds lightning crashes.

Tape stop and call drop

Tape stop cuts motor power so the read head decelerates along a (1 − t)¹·⁶⁵ braking curve to a dead halt while rolling off highs, ending with a relay clunk. Call drop cuts the line dead with a click, silence, and the North American 480 + 620 Hz fast busy signal.

Speaker and space

Each speaker is a chain of band filters, compression and soft clipping. Telephone keeps 300 Hz to 3.4 kHz and AM radio about 200 Hz to 4.5 kHz. Room is a convolution reverb and Echo is a filtered feedback delay.