[bdbarnett]

A sampler voice can now play the rest of what a SoundFont region says about a note:

• a two-pole resonant low-pass (cutoff in Hz, resonance in dB),
• a DAHDSR modulation envelope driving pitch and cutoff (mod_*),
• a vibrato LFO on pitch (vibrato_delay, vibrato_rate, vibrato_depth),
• a modulation LFO on pitch, cutoff and volume (lfo_*).

They're new keywords on play(), on CPython, MicroPython and CircuitPython. cutoff and vibrato_depth are also live properties, so a mod wheel or a velocity curve can move them on a sounding note.

How it renders

A voice that sets none of these takes the old path and renders the same bits; sampler_probe.py's output matched main's before its new section was added. A voice that does set them renders in 32-frame chunks. Once a chunk it works out its pitch, its filter coefficients and its volume, then runs the chunk through a state-variable low-pass. The coefficients come from a polynomial tangent and audiodsp_exp2_poly, so CPython and MicroPython print the same probe output.

What it costs

tests/parity/sampler_cost.py gives microseconds per 256-frame stereo block at 48 kHz, where the block lasts 5333 µs. filter is a 2 kHz low-pass with 6 dB of resonance. full adds the modulation envelope and both LFOs on everything they reach.

Voices	CPython plain / filter / full	ESP32-P4 plain before / plain / filter / full
32	32 / 56 / 65	1383 / 1384 / 2626 / 3465
48	47 / 83 / 95	2039 / 2040 / 3904 / 5179
64	62 / 102 / 122	2745 / 2804 / 5290 / 6970

On the P4, plain before is the same firmware built from main. At 64 voices the plain path reads 2% slower, which is about the spread between two images of the same code: main's sampler read 2704 µs at 64 voices in a bigger image. The voice struct grows from 112 to 256 bytes on a 64-bit desktop. The numbers are in docs/how-it-works/upstream-diff.md.

Tests

tests/test_cpython_audiosampler_modulation.py has ten traits, each with its bar:

• An open filter renders the same bits as no filter.
• The response at and below the cutoff is within 0.3 dB of the design, and two octaves up it is 21 to 27 dB down.
• The envelope sweeps pitch and cutoff.
• The LFOs' rate, depth and delay are measured.
• The two live properties take effect.
• 32, 48 and 64 filtered voices fit inside a block.

Three planted faults each failed tests:

• resonance at half its dB failed the response trait,
• a triangle at twice its phase failed the LFO traits,
• the envelope at half speed failed the sweep.

The CircuitPython binding is written but hasn't been built here.