audiodsp

CircuitPython's audio system for MicroPython and CPython — audiocore, audiomixer, synthio (including MidiTrack), the effects modules (audiofilters, audiodelays, audiofreeverb, audiospeed), audiomp3, and CircuitPython play()/stop()/pause()/resume() output semantics. Import names stay audiocore/synthio/etc., matching CircuitPython for source compatibility.

audiodsp was named audioif until 2026-09-21; that name now belongs to the audio hardware layer. The modules you import did not change. The helper package is audiodsp_util now, and the distribution on PyPI is pydevices-audiodsp.

New here? See the newcomer's guide for the runtime map, shared-core architecture, and contributor boundaries.

Installation

MicroPython 1.29 or later takes this repository with one line in the manifest your build already uses. Clone it anywhere, fetch its two pinned native dependencies, and include its manifest.py:

git clone https://github.com/PyDevices/audiodsp ~/build/audiodsp
~/build/audiodsp/scripts/fetch_deps.sh

include("/home/you/build/audiodsp/manifest.py")

On unix the manifest to edit is ports/unix/variants/standard/manifest.py; on esp32 and rp2 it is usually ports/<port>/boards/manifest.py, unless your board brings its own. Then build the way you always do. The include brings in the audiodsp C modules and ulab (named with c_module()); it freezes no Python. fetch_deps.sh puts ulab and mp3 under .deps/, pinned by DEPENDENCIES.lock; the manifest and the mp3 glue look there first and fall back to a sibling checkout beside this repo. Tested on the unix port against MicroPython v1.29.0, where synthio, audiomp3, audiopump and ulab.numpy all import.

If you would rather not edit the MicroPython tree, write a manifest of your own and pass it as FROZEN_MANIFEST=. That replaces the port's default, so include the default too (include("$(PORT_DIR)/variants/standard/manifest.py") on unix, include("$(PORT_DIR)/boards/manifest.py") on esp32) or you lose asyncio and the port's other frozen modules.

To build without mp3, skip the fetch and set AUDIODSP_OPTIONAL_DEPS=1 on the build step — micropython.mk/micropython.cmake read it, not fetch_deps.sh, and on CMake ports it is an environment variable, not a -D cache entry.

Older than 1.29? Manifests there have no c_module(), so do not include this one; use USER_C_MODULES. On esp32 and rp2 point it at this checkout (add ;<checkout>/.deps/ulab/code if you want ulab); on a Make port point it at the directory that contains the checkout, which builds every module in that directory and does not reach ulab. Without ulab, synthtools' import ulab.numpy fails at runtime.

MicroPython's manifest reference explains include() and c_module(), and micropython-pydevices keeps ready-made manifests and boards for several PyDevices modules at once. docs/porting-plan.md covers the architecture, module tiers and testing strategy.

CPython 3.11+ installs from PyPI:

python -m pip install pydevices-audiodsp

This gets you audiocore, synthio, audiomixer, audiofilters, audiodelays, audiofreeverb, audiospeed, audiodynamics, audioroute, audiomath, audioecho, audioshaper, audioladder, audioconvolve, audiobiquad, audioverb, audiomodal, and the audiorender package. audiomp3 remains firmware-only. The distribution declares no required runtime dependencies and does not itself publish an audiodsp import; its version is the VERSION file, which is also what _audiodsp.__version__ reports.

audiorender is the exception, and it is opt-in: it is numpy throughout, so numpy comes with the render extra rather than with the wheel, keeping the core dependency-free for boards and wasm:

python -m pip install "pydevices-audiodsp[render]"

The instrument and effect libraries — audioinstruments (55 synthesizers, keyboards and drum machines) and audioeffects (45 effect classes, racks included) — are not part of this distribution. They live in the audiocomponents repository as their own packages, each depending on pydevices-audiodsp; see that repository for how to install them.

What's here

Two pure-Python modules sit beside the CircuitPython-compatible core:

• lib/audiodsp_util/ — float32(value) and float32_bits(value), struct and nothing else. A Python float is the interpreter's own width: a double here, a single on every board. So a setting written in Python — node.mix = 0.35 — is two different numbers, and the node renders different bytes on a board than on a desktop before its kernel is reached. Put a derived setting through float32 and both targets hold the number the board would have held. See docs/correctness-standard.md.
• lib/audiorender/ — renders a whole composition offline: tracks, tempo map, notes and automation in, a mixed stereo master and a level report out. This is the one part of the repository written for a desktop rather than a board — numpy throughout, a whole song in memory — so it ships in the wheel and is never frozen into firmware. It has its own README.

The instrument and effect libraries that used to sit beside it — audioinstruments and audioeffects — now live in the audiocomponents repository, together with the audio component contract they implement: the metadata manifest (docs/audio-components.md) and the runtime API (docs/audio-component-api.md). audiorender drives any component that speaks that API and does not need those packages installed.

Additions beyond CircuitPython

Ten things here are not CircuitPython's. audiodynamics (compression, limiting, expansion, gating, transient shaping) and audioroute.Splitter (fan one stream out to parallel branches) come from micropython-vst3's audio engine, which had them and CircuitPython does not. audiomath (multiply one stream by another — ring and amplitude modulation; and divide one down in frequency — the analog octave divider), audioecho (a delay with a filter, a soft-clip and a cross-feed inside its feedback loop), audioshaper (a waveshaper whose curve is data, applied above the sample rate, and a sample-and-hold at an exact rational ratio), audioladder (a transistor ladder filter — four one-pole stages round a feedback loop with an odd saturator inside it), audioconvolve (apply a measured or synthesized impulse response, by partitioned FFT), audioverb (a reverberation tank whose line lengths and output taps come from Python), audiomodal (a bank of resonators, which is what a struck object is), audioroute.MidSide (scale the difference between a stereo pair's channels) and audiobiquad (below) have no ancestor anywhere and are audiodsp's own. apply_cp_patches.sh adds every one of them to a CircuitPython tree too.

audiobiquad — filters whose tails reach exact zero

audiofilters.Filter (over synthio.Biquad) and audiofilters.Phaser are ported CircuitPython and their recursions are integer. Both have states that reproduce themselves: fed silence after a burst, they hold a constant for ever. audiobiquad is the same two filters with float state and a flush of anything below 1e-20, so a tail decays to zero and stays there — and its all-pass feedback is not clamped to 0.1..0.9, so zero is zero and a feedback-free phaser's notches are true nulls. It is also the one to reach for below about 100 Hz, where the integer coefficients run out of resolution — which filter, and why.

import audiobiquad, synthio

eq = audiobiquad.Biquad(mode=audiobiquad.PEAKING_EQ, frequency=3200,
                        Q=1.2, gain_db=-4.0, sample_rate=48000)
eq.play(source)

sweep = synthio.LFO(rate=0.4, scale=400.0, offset=900.0)
phase = audiobiquad.AllPass(stages=4, frequency=sweep, feedback=0.0,
                            mix=0.5, sample_rate=48000)
phase.play(eq)

mode is one of LOW_PASS, HIGH_PASS, BAND_PASS, NOTCH, PEAKING_EQ, LOW_SHELF, HIGH_SHELF — numbered exactly as synthio.FilterMode numbers them, so either can be handed to either. mix crossfades: 0 is a wire, 1 is the filtered signal alone; for an all-pass cascade 0.5 is the equal sum a script phaser makes, and where its notches are deepest. AllPass's frequency really is the frequency at which one stage's phase passes −90°, so nothing has to pre-warp it. Both nodes hand out 256 frames at a time, sit in an audiosample chain like any other effect, and never report themselves finished — a starved chain gets silence. Q is bounded to 0.05..60 and feedback to ±0.99: a pole on the unit circle is a filter that never stops ringing, which is the one thing this module exists to avoid. See docs/upstream-diff.md for the measurements.

audiodynamics.Dynamics carries twenty-one options beyond what the engine gave it, added for the effects program and every one of them default-off: an RMS detector, a feedback detector topology and an external key input with Key Listen, a 4x true-peak reconstruction, a two-pole key band with both corners settable, a settable expander/gate depth, a four-stage gate envelope with hold and hysteresis, a relative-threshold gain computer, a program-dependent attack, and the transient shaper's four detector time constants, a second envelope pair and a peak-hold on the slow one. What each is for, what it was measured doing, and what it cost: docs/upstream-diff.md.

audioecho.FeedbackDelay takes four further options, each off at its default. wow_shape swaps the built-in modulation sine for one period of your own — int16 Q15, a power-of-two count from 2 to 4096 — because a bucket brigade's delay is its line length over its clock, so a triangle on the clock is a reciprocal on the delay and no sine is that. delay_slew walks the read head to a new delay_ms at a constant rate (delay-seconds per second) instead of jumping to it, which is both an Echoplex's varispeed and the reason a delay-time knob can be turned mid-take without a click. wow_am_depth puts the same oscillator on the wet gain, dipping only. loop_semitones pitch-shifts the line read inside the loop, so every repeat rises again — a shimmer, which chaining audiodelays.PitchShift after a delay is not. See docs/upstream-diff.md for what each was asked for and what it measures.

audioroute.MidSide(source, width=1.0, sample_rate=48000, channel_count=2) takes a stereo pair apart into its mono sum and the difference between its channels, scales the difference, and puts the pair back together. width=0 collapses to mono, 1 passes through, 2 doubles the sides; values outside 0..2 clamp to those rails. play(sample) sets the source, set(width=...) moves the width mid-stream. There is no state and no latency — and at width=1 the output bytes are the input bytes, exactly, for every int16 pair, so the node costs nothing to leave in a chain that is not using it. A mono source passes through: there is no difference to scale. It exists for the drive classes as much as for stereo width: a nonlinearity applied to a stereo pair intermodulates its channels, so a saturator that wants to keep its image drives the mid and leaves the side alone.

audioshaper.Waveshaper(sample_rate=…, curve=…, oversample=4, channel_count=2, **options) is the newest of them, and everything but the first four moves in set(). curve is int16 Q15, at least two points, spanning −1..+1 of input: compute it once on a desktop and ship it as data, never rebuild it on a board, whose float is single-precision where the desktop's is double. oversample is 1, 2, 4 or 8 — the shaping happens there, between a matched pair of polyphase all-pass half-bands, because a nonlinearity at the base rate folds the harmonics it makes above Nyquist straight back onto the signal. pre_gain is the drive knob (gain into one normalised curve, never a curve rebuilt per knob move), bias moves the operating point, post_gain follows the curve, and mix is a straight 0..1 crossfade. hysteresis is off by default; above zero it gives the curve a memory, so a slow signal in and out traces two paths and encloses an area, with hysteresis_width its half-width as a fraction of full scale at the input and hysteresis_bias splitting that between the rising and falling branches. audioshaper.GROUP_DELAY_SAMPLES carries the measured group delay per factor — 0, 2.2, 3.3 and 3.9 base samples — for a component that has to report its latency. A curve that reaches the rails has a ceiling of its own: the decimator can ring past them on a hard edge, so keep post_gain * max(abs(curve)) at or below audioshaper.CLIP_HEADROOM (0.74) and put the rest of the wanted level on a mixer voice after this node. Why each of those is the shape it is, and what the oversampling measures: docs/upstream-diff.md, "audioshaper: audiodsp's own, and the two things a fixed curve cannot be".

audioshaper.SampleHold(source, num=…, den=…) is the same module's other node: a zero-order hold in which num source frames carry den new values, so 26 040 Hz at 48 kHz is 400/217 and is exact — one frame in, one frame out, at the source's own rate, channels and bit depth, with an integer accumulator that cannot drift however long the render runs. num == den is a wire byte for byte, latency is 0 (what a hold displaces is an event landing on a frame it drops, up to ceil(num/den) − 1 frames, which is the effect and the caller's to report), and set(num, den) moves the ratio mid-stream. It exists because the palette's sample-and-hold was a pair of audiospeed.SpeedChanger nodes whose 16.16 rates cannot be made reciprocal: measured on that hold, the pair ran 0.9999947 of the rate it asked for and flanged a static setting by 10.74 dB over twelve seconds. The rate form was not added to SpeedChanger because that module is CircuitPython's and an argument added here would not exist on a board — docs/upstream-diff.md, "audioshaper.SampleHold: lo-fi's other half".

audioladder is the newest, and the one whose reason for existing is least obvious next to a module CircuitPython already has. audiofilters.Filter is a better resonant low-pass than this: a cascade of biquads tracks the analytic response to a fraction of a decibel. It is also linear, so it never sustains a tone of its own, has no harmonics to speak of, and sounds exactly the same however hard you hit it — and those three are what a ladder is. The whole surface:

import audioladder

ladder = audioladder.Ladder(
    sample_rate=48000,      # and channel_count=1 or 2
    cutoff_hz=800.0,        # where the filter turns over
    resonance=3.6,          # 0..4.2; at 4 it sustains a tone at the cutoff
    drive=2.5,              # linear gain into the loop, 1 being unity
    poles=4,                # tap stage 1..4: 6, 12, 18 or 24 dB an octave
    passband_comp=0.0,      # 0..1, how much of the passband to give back
    oversample=2,           # 1 or 2; 2 folds back fewer of its harmonics
    mix=1.0,                # 0 a wire, 1 the filter
)
ladder.play(source)
ladder.set(cutoff_hz=1200.0)   # any option, mid-stream, once a block
ladder.clear()                 # empty the integrators; stops a sustained tone

See docs/upstream-diff.md for why its feedback loop is solved rather than delayed, and what that was measured to be worth.

audiomodal.Bank

A bank of resonators, which is what a struck object is. Hit a drum head, a marimba bar, a bell or a wine glass and it rings as a sum of decaying sinusoids at frequencies that are not harmonics of anything; hand Bank that list and a short noise burst for the stick, and it is that object.

kick = audiomodal.Bank(modes=9, sample_rate=48000, channel_count=1)
kick.set_modes(((58.0, 0.55, 1.00), (92.4, 0.24, 0.38),
                (123.9, 0.15, 0.24), (2200.0, 0.010, 0.30)))
kick.play(stick)          # a short noise burst
audio_out.play(kick)

Construction fixes the allocation: sample_rate, channel_count (1 or 2) and modes (1 to 64). set_mode(index, frequency, decay, gain) places one mode and set_modes(table) places the whole list; decay is a 60 dB time in seconds, not a Q, and gain is the peak of that mode's impulse response, so a modal table read out of a paper goes in as published. set() moves mix and gain mid-stream without stopping anything ringing, clear() stops everything at once, and ringing says whether any mode still holds energy.

Two behaviours worth knowing before you use it. A bare Bank() is silence, not a passthrough — the default mix is fully wet and an unconfigured bank has no modes — so reach for mix=0.0 if you want the untouched signal. And unlike audioecho.FeedbackDelay and audioverb.Tank, the tail keeps ringing after the source stops, because a drum whose decay ended the instant the stick left would be the one thing this node exists not to be.

See docs/upstream-diff.md for why this is a node rather than a chain of audiobiquad.Biquad, measured rather than argued.

audioverb.Tank

Dattorro's plate reverberator, with the network handed in rather than compiled in — which is the difference from audiofreeverb.Freeverb, whose comb and all-pass lengths are constants in a CircuitPython-ported kernel.

plate = audioverb.Tank(
    sample_rate=48000, channel_count=2, max_predelay_ms=200,
    decay=0.7, diffusion=0.75, bandwidth_hz=9000, damping_hz=4500,
    mod_rate_hz=1.0, mod_depth_ms=0.27, mix=0.35)
plate.play(source)
audio_out.play(plate)

Construction fixes the allocation and the topology: sample_rate, channel_count (1 or 2), max_predelay_ms, delays and taps. delays is twelve line lengths in frames — the four input diffusers, then each tank half's modulated all-pass, delay, all-pass and delay. taps is four values per tap (channel, line index, offset in frames, gain), at most 32. Both default to Dattorro's published table, scaled from 29761 Hz to sample_rate.

set() moves the rest, mid-stream and without emptying the lines: decay, diffusion, damping_hz, bandwidth_hz, low_cut_hz, predelay_ms, mod_depth_ms, mod_rate_hz, drive, width, tone_db, mix. Every filter is out of the path at zero, so a bare Tank() is that network with nothing added to it; mix follows audiodelays.Echo's 0..2 convention, dry at unity until 1. clear() empties every line and every filter. Latency is zero dry-to-wet, and the tail reaches exact zero rather than sitting at one LSB — every line write is a magnitude truncation, which is the standard cure for a recirculating fixed-point network's limit cycles. See docs/upstream-diff.md for the measurement, the cost and the memory the default table needs.

audioroute.Port — the wire a consumer holds

A zero-copy pass-through you can re-point while it is playing. Its get_buffer hands back its source's own pointer, length and result; its play(source) swaps that source atomically against a pull in flight. Hold a port, wire it into a mixer voice or a rack, and change what is behind it as often as you like — the object the consumer took never changes.

import audioroute

port = audioroute.Port(overdrive)
mixer.voice[0].play(port)       # take it once

port.play(bitcrusher)           # heard on the next block; nothing rewires
port.source                     # what it is playing right now

That is the problem it was built for. About twenty effect classes rebuild part of their graph when a knob crosses a threshold and hand out a new node afterwards — and a consumer holding the old one goes on playing the graph the class has finished with, or goes silent. A port turns that into a re-point. play() does the protocol lookup, the deinit check and the format match first, on the calling thread where raising is free, then takes the pump lock for three stores, so a pull in flight sees the whole old source or the whole new one and nothing in between.

It costs about 20 ns a block on x86-64 and 1.27 µs a block on an ESP32-P4, both measured as the slope of 0, 25, 50 and 100 stacked ports over the same 2000-block probe; against a 256-frame block at 48 kHz that is four ten-thousandths of one per cent on the desktop and about two hundredths of one per cent on the board.

The board figure replaces an estimate of 700 ns that was extrapolated from four effect classes timed on both machines, and the estimate was low by 1.8x. The real ratio here is 64x, not the 32–35x those classes gave: a port's work is a protocol lookup and three stores with almost no arithmetic, so it does not scale the way a DSP node does. That is an argument against extrapolating a per-class cost at all (#113).

Port has no stop(), deliberately — a port always has a source. A port pulled while it is already inside itself, which is what a component wrapping its own output builds, publishes a loop fault and hands back an error in 17 us instead of recursing until the stack is gone.

audiopump — the audio pull, off the interpreter thread

audiopump runs the same pull a player runs, in C, on a thread the interpreter is not on: a task pinned to the other core on esp32, a native thread on unix and Windows, and the calling thread on a port that has no threads at all. A graph pulled that way keeps an exact clock while a screen redraws, a USB stack runs and Python does whatever it likes.

import audiopump, audiomixer, synthio

synth = synthio.Synthesizer(sample_rate=48000, channel_count=2)
mixer = audiomixer.Mixer(voice_count=1, sample_rate=48000,
                         channel_count=2, buffer_size=2048)
mixer.voice[0].play(synth)

status = bytearray(audiopump.STATUS_BYTES)
audiopump.spawn(mixer, 0x7FFFFFFF, status, sink=True)

synth.press(60)                # while it plays; no parking, no ceremony
mixer.voice[0].level = 0.6
audiopump.shutdown()           # and a soft reset does this for you

spawn(sample, blocks, status, …) starts it; pull() runs the same loop on the calling thread, so micropython.heap_lock() around it is a real gate on "the pull does not allocate"; service() advances it where there is no thread; retarget(sample, loop=…) points it at a different tail; shutdown() stops it. status is a bytearray of counters — blocks, bytes, a digest of everything pulled, the worst block, what the sink clocked, the error and the fault — read back with struct.unpack. Everything that can refuse happens on the calling thread. The pull itself never raises, because there is no interpreter on that thread to raise on, so it publishes a fault code and stops.

retarget's loop= travels with the swap rather than being stored when you call it: the old tail is pulled until the next block boundary, so setting the flag early ends the pump one block before the new tail ever runs. Leave it out and the flag stays as spawn() set it, which is what every caller before the argument existed wanted. Get it wrong the other way and a looping client left alone on a live pump stops at the end of its lap.

backpressure() — whether a full ring makes the pump wait

>>> audiopump.backpressure()
True

True where a full output ring makes the pump wait for room instead of dropping the block; False where a free-running pump would lose audio. It is the question a driver asks before it decides whether to park the pump between ticks, and it is True in two different ways: on a threaded port whose driver fills in park_spin, where a wake ends the wait, and in service mode, where the loop hands the thread back on a full ring and the caller's next service() is the wake. It is False only on a threaded build with no park_spin — a driver that is not finished, which now says so rather than quietly losing blocks.

It is worth asking because parking costs the caller. On the desktop unix build, ten seconds of a synth through an Overdrive and a TapeDelay with an app doing 3.5 ms of work a tick: the interpreter spends 52 ms per 10 s on audio against 862 ms on the old interpreter-thread path, where a parked pump cost 923 ms. With no app work at all the two are level (758 ms against 713). A parked pump used to spin 94 % of one core and now sleeps at 1 %. A WAV plays back byte-identically ten times out of ten with all eight cores of that box in a busy loop, overflow count 0 by construction; with the drop put back, 0 of 10.

Three kinds of source feed it and they are all the same kind of thing to it: a graph; audiopump.Ring, an audiosample node Python writes PCM into, so a pushed stream can sit behind a Mixer with effects on it like anything else; and audiopump.Events, a queue of frame-stamped presses, plays and levels the pump applies at block boundaries, which is what puts a sequenced bar on the audio's clock instead of the interpreter's. audiopump.Tap reads what is going out, for a meter or a scope, without being in the path.

Which builds carry it. Every MicroPython port: unix, Windows, WebAssembly and esp32. The CPython wheel does not — it is a MicroPython C module through and through (MP_REGISTER_MODULE, mp_obj_t, a VM root array), so it is excluded from the wheel rather than stubbed, and a desktop Python program has threads of its own. CircuitPython does not either: its playback layer already pulls natively from its own audio thread, and a second pull loop would be a second owner of the same graph. What both of those still take is the lock and the hook table, on default hooks, which cost a load and a branch.

The hardware — the channel a board writes into, and a live microphone as a source — is not here. It arrives as a separate module, _audioif, from a platform driver, and audiopump.driver() tells you which one bound:

>>> audiopump.driver()
'esp32'          # or 'pthread', 'win32', 'none'

'none' is a real answer rather than a failure: the default hook table is all NULLs, which means one thread and no hardware, and that is exactly how the WebAssembly build runs. Writing a driver for a platform that has none yet is docs/pump-ports.md.

Status

MicroPython: all module tiers ported and oracle-diffed byte-for-byte against bin/circuitpython on unix; DSP parity re-verified on windows and wasm; built and measured on two real mcu targets, ESP32-P4 (hardware-confirmed by ear) and RP2040 (build-only). See docs/porting-plan.md for the full phased history and docs/upstream-diff.md for every deliberate deviation from upstream CircuitPython.

CPython: the public surface and wheel plumbing are present, and the committed synthesis, mixer, MIDI, streaming, and effects fixtures match CircuitPython 10.3.0 PCM byte-for-byte — built at the same voice ceiling this port ships, which is the only comparison worth making. Import/API smoke success is not used as a substitute for those comparisons.

ulab and mp3 (the vendored Adafruit_MP3/Helix decoder audiomp3 depends on, RPSL/RCSL-licensed — not MIT, carried unmodified per upstream's own terms) are consumed as cloned sibling dependencies in the parent workspace, same pattern as pygraphics/displayif, not vendored into this repo.

The playback-facing pull protocol (audiocore.get_buffer/reset_buffer) is consumed by pydevices' lib/audiodev package (AudioOut in sample_out.py), which pumps any audiosample into any of audiodev's existing push-PCM transports (sdl2/win/wasm/i2s/emulated) — see that package's own docs for the playback-side contract this repo's protocol implementation is built to satisfy.

Acceptance target: todbot's synthtools running unmodified, with rendered PCM diffed against bin/circuitpython's own unix coverage build (which already contains the entire DSP stack — the parity oracle for this whole port).

Sound stability

The API is our contract with you: class names, signatures, metadata, and macro surfaces stay stable and change only deliberately. The sound is not part of that contract. These components sound great, but they are not all as accurate as they could be, and implementations will keep being refined as the library matures — so a component may render audibly differently from one release to the next. If a composition depends on the exact sound of a release, pin that release rather than tracking the latest; the code of every release stays available for exactly this reason.

Beneath the components sits a harder guarantee: the audiodsp core — the CircuitPython-compatible synthio/audiocore/effects-module layer — is held bit-exact to CircuitPython itself, verified by parity gates, and that never changes release to release. Where we find CircuitPython and audiodsp disagree, we treat it as a bug and report it upstream. The components are where the sound evolves; the floor they stand on does not.

License

audiodsp is MIT licensed — see LICENSE.

The attribution that goes with it lives in NOTICE: substantial portions of this code are ported from CircuitPython under its own MIT license, CircuitPython is a trademark of Adafruit Industries and this project is not affiliated with or endorsed by them, and the vendored test corpus under tests/vendor/ carries its own licenses. The two are separate files on purpose — GitHub and PyPI only detect the license when LICENSE holds the MIT text and nothing else, so please do not fold NOTICE back into it. NOTICE is in setuptools' default license-files glob, so it ships in the wheel and sdist alongside LICENSE.

The Helix MP3 decoder that audiomp3 wraps is RPSL/RCSL-licensed, not MIT; it is a cloned sibling dependency, not part of this repository. See docs/upstream-diff.md.