Play FM voices

A bell voice from audiofm: two struck notes ringing out, their inharmonic partials in the spectrogram, and 40 ms of the waveform

Six sine operators per voice, each with its own envelope, wired into one of 32 algorithms where some operators modulate the frequency of others: that is the FM synthesis of the classic six-operator instruments, and audiofm plays their voice format. Bring a bank of 32 voices, a .syx file of 4104 bytes, and play any voice in it.

import audiofm

with open("bank.syx", "rb") as f:
    bank = audiofm.Patch.bank(f.read())
synth = audiofm.Synth(voices=16, sample_rate=48000, patch=bank[0])
audio_out.play(synth)
synth.press(60, 100)        # note, velocity
...
synth.release(60)

A Synth is an audio source like synthio.Synthesizer: play it into an output, a mixer voice or an effect. press() takes a free voice, or steals the oldest released one and then the oldest held one when all of them are sounding. sustain, pitch_bend (with bend_range in semitones) and mod_wheel are the keyboard’s controls. The mod wheel adds LFO vibrato and tremolo on top of the voice’s own.

Edit a voice

Every parameter is an attribute, in the format’s own whole numbers, so a value read off a voice chart goes straight in. operators[0] is operator 1.

patch = audiofm.Patch()          # the initial voice: operator 1, a sine
patch.algorithm = 5              # three modulator and carrier pairs
patch.feedback = 6
op = patch.operators[1]          # operator 2 modulates operator 1
op.level = 72
op.coarse = 1                    # ratio 1; 0 is one half
op.rates = (95, 50, 35, 78)
op.levels = (99, 75, 0, 0)
synth.patch = patch

synth.patch is read at each press(), so an edit reaches the next note and never one already sounding. A value outside its range raises ValueError. operators[n].enabled = False mutes an operator without losing its settings.

Save and load

patch.pack() gives the 155-byte voice and Patch.unpack() reads it back, or reads a 163-byte single-voice message with its checksum. Patch.pack_bank(patches) writes up to 32 voices as a 4104-byte bank, and Patch.bank() reads one, checking the checksum. Values past a parameter’s range in a file are clamped rather than refused, so a bank from an old editor still loads.

with open("mine.syx", "wb") as f:
    f.write(audiofm.Patch.pack_bank(bank))

How loud it is

One note of a loud voice peaks at a sixteenth of full scale or less. That leaves room for sixteen notes of the loudest voices at once without clipping. If you play one line at a time and want it louder, raise it in a mixer.

What it costs

The engine is fixed point, and an operator whose level is below a tenth of an LSB isn’t computed, so a voice costs what is sounding in it. A released voice frees itself once every carrier has fallen silent. Rendering is identical on CPython and MicroPython, which tests/parity/fm_probe.py checks.

The engine is adapted from Music Synthesizer for Android; what came from where is in third_party/msfa/ORIGIN.txt.