SIREN / design lab / parameter reactivity
Tell the chip where to go — and how to get there. Spend your updates on the strike, then let the note breathe between instructions.
Chipset proposalInteractive curvesExisting piano data
The CPU supplies landmarks. The chip draws the journey. Gain, pan, send, and other continuous controls can move smoothly even when their next instruction arrives only ten times a second.
New: tune bits, point count, and CPU playback rate independently →
The compression lab packs real descriptors and explores a 600 Hz CPU interpreter
without changing the chip.
Loading the current piano curves…
01 / Separate instructions from movement
A new value becomes active at the next control latch. No journey — just the destination.
Useful for deliberate steps, switches, and resetting a value. Sparse gain updates sound stepped unless an envelope supplies the motion elsewhere.
Each audio sample closes a fraction of the remaining distance. New targets can arrive whenever they like.
T₉₀ is the time to cover 90% of the gap. Calculate α only when the response time changes; the sample loop needs one multiply and two adds/subtracts. It glides naturally, but follows behind a moving target.
“Reach this value in 25 ms.” The player sends the next landmark at the start of the segment.
One addition per sample, plus a countdown. Exact arrival time. Perfect for planned envelope landmarks; also useful for a pan sweep or send fade.
The same scheduled arrival, with a gentle departure and landing. No overshoot between two points.
Smoothstep has zero slope at both ends. In gain space, its cubic can run with forward differences — three additions per sample after setup. Log-space motion also needs conversion back to gain. Rounding every landmark can sound too cushioned; the linear option stays useful.
Two different instructions: “follow this target” reacts to the present; “arrive there in this time” plans the next segment. For a piano envelope we already know the next point, so timed ramps can avoid the follower's lag.
02 / Put the idea on the actual curves
The white trace is the current 440-reduced piano's gain curve. The dots are the landmarks kept by the selected schedule. All motion modes share those landmarks: the follower receives each value when its dot arrives; timed ramps receive the next value one segment earlier.
The count is the number of intervals after initialization at t = 0. Each interval corresponds to one set of eight target values: ramp commands go out at its start, follower/step commands at its end. A schedule determines when to send targets, not how many bytes describe the instrument.
| Time | Gap from previous | Gain / peak | dB | Ramp command sent at |
|---|
For a follower or immediate mode, send at the landmark time. For a ramp, send at the preceding landmark with the gap as its duration. The same coefficient may be reused while response settings stay unchanged.
The reference comes from the existing 240 Hz gain table, joined linearly for this drawing. A 600 Hz view explores scheduling and interpolation between those values; new attack measurements can later supply finer detail. The follower always operates in linear gain; “ramp geometry” changes only the two planned ramps. No audio plays on this page.
03 / A clock that stretches with the note
Pure doubling — 1, 2, 4, 8, 16… — quickly leaves very large gaps. A gentler geometric progression keeps more of the broad dips and recoveries.
At first, move one chip tick at a time. Then the gap grows with note age, roughly multiplying the time by 9/8 per update. Eventually cap the gap at 60 ticks: 100 ms, or 10 Hz. The integer version lands on the chip's existing 600 Hz latch grid.
| Note age | Interval | Equivalent rate |
|---|---|---|
| 0–10 ms | 1.67 ms | 600 Hz |
| 20 ms | 3.33 ms | 300 Hz |
| 60 ms | 8.33 ms | 120 Hz |
| 200 ms | 25 ms | 40 Hz |
| 800 ms and later | 100 ms | 10 Hz |
Check “keep a 600 Hz strike” to give the first 20 ms a uniform fine grid. After that, the expanding rule or chosen slower schedule takes over. Knots are a budget, not a prison: a known sharp change can request an extra landmark wherever it is needed.
Leave the global latch at 600 Hz. A quiet tail doesn't need to write every tick, but a new note can still start promptly. Each parameter's current value continues evolving at audio rate between latches.
For actual CPU savings, eventually schedule the next event rather than run a high-frequency interrupt that mostly does nothing. Multiple notes can share one queue of upcoming landmarks. Their attacks and tails naturally use different cadences.
04 / A small language for moving parameters
My suggested first interface is per-parameter policy, initially for the eight slots' gain, pan, and send: 24 independent continuous lanes. Keep source selection and routing as discrete choices.
This is the proposed command vocabulary; register addresses come after choosing the behavior. Configuration persists, so repeated targets need not repeat their mode and time.
A new ramp starts from the value the chip is currently using, not from the previous command's destination. An exponential follower keeps its current value and changes its destination. Immediate mode jumps deliberately.
Rounded ramps restart with zero slope; value is continuous, velocity can change. A future slope-preserving cubic is an option if that corner matters.
Defaults remain immediate, so existing programs keep their behavior. Mode/time and target changes latch together. Rewriting an unchanged target does not restart a ramp; a separate retrigger command can request that.
CPU reads continue to return programmed targets. Capture/debug tools can expose the evolving current values. These are two different things, and both are useful.
Timed ramps snap to their endpoint. Exponential followers settle to an exact target below a small internal threshold instead of leaving endless floating-point tails. A zero time means immediate.
Clamp gain/send to their valid range and pan to its signed range; reject invalid commands. All low-level motion uses fractional internal precision, rather than rounding to a register byte every sample.
Gain-space ramps can start and end at zero naturally. A straight line in dB produces an exponential amplitude ramp. Musical pitch glides should move in semitones; pan moves in its signed coordinate.
Log-gain plots use an −80 dB floor between zero endpoints. The engine can implement positive log ramps as repeated multiplication, with explicit start/stop handling for zero. Filters, pitch, and delay feedback deserve their own domain and stability rules before joining the first 24 lanes.
Recommended first implementation: immediate, exponential following, and timed linear ramps for gain/pan/send. Keep the rounded mode in the playground while listening decides whether its extra curve state is useful. Keep native generator envelopes for attacks they already express well.
05 / Small descriptions, small moving state
A four-byte partial description can generate a destination only when the next landmark is due. The chip keeps the current value, destination, and a little motion state — it does not need an eight-second expansion of all future values.
A simple fixed allocation could reserve eight 32-bit state words per lane: 24 × 32 = 768 bytes, plus mode/configuration storage. Linear and follower modes can use less. This is internal chip state, separate from the compact instrument bytes.
The schedule can be generated from the rule; it need not be stored per partial. Shared attack/tail settings and the small decoder still count in the instrument/player budget.
Only moving lanes need updates. A follower uses a multiply and two adds/subtracts per sample; a linear ramp needs an add and countdown. Setup performs divisions or exponentials once per new motion, not 48,000 times a second.
At most 24 first-stage lanes × 48,000 samples = 1,152,000 lane steps per second. Guest control traffic falls; host interpolation work rises. We'll measure both when implementing it.
There are two independent compression wins: compact shapes reduce instrument data, and sparse destinations with chip-side motion reduce control traffic. Together they let us keep organic changes without paying for thousands of nearly repeated gain words.