Skip to content

Latest commit

 

History

History
822 lines (568 loc) · 30 KB

File metadata and controls

822 lines (568 loc) · 30 KB

Audiophile-grade DSP processing on Apple Silicon

Prepared for: C++23 + Swift music player design
Verified on: 2026-06-17
Scope: macOS music playback, digital signal processing, Apple Silicon acceleration, open-source audio projects, and AI-based enhancement.


1. Executive conclusion

The best architecture is a hybrid stack.

Swift / SwiftUI
→ AVAudioEngine / Core Audio output
→ C++23 real-time DSP kernel
→ Accelerate / vDSP for filters, FFT, convolution
→ optional libebur128, soxr, FFmpeg
→ optional Core ML / MLX for offline AI features

Use Apple-native APIs for playback, output routing, device negotiation, and app integration.

Use C++23 for the real-time digital signal processing kernel.

Use selected open-source libraries only where they provide measurable value.

Do not make AI enhancement the default path.

Use AI for offline restoration, stem separation, denoise preview, tagging, and low-quality recording enhancement.

Use classical digital signal processing for equalizer, loudness normalization, limiter, sample-rate conversion, room correction, crossfeed, and true-peak control.


2. Definitions

DSP means digital signal processing. It is math applied to audio samples.

PCM means pulse-code modulation. It is the common raw digital audio format.

FFT means Fast Fourier Transform. It converts audio from time view to frequency view.

FIR means finite impulse response. It is a filter without feedback.

IIR means infinite impulse response. It is a filter with feedback.

dBFS means decibels relative to full digital scale. 0 dBFS is the digital clipping point.

dBTP means decibels true peak. It estimates peaks between digital samples.

LUFS means loudness units relative to full scale. It measures perceived loudness.

Core ML is Apple’s machine learning framework.

MLX is Apple’s open-source array framework for machine learning on Apple Silicon.


3. What “audiophile-grade” should mean

“Audiophile-grade” is not a formal engineering standard.

Use measurable transparency instead.

Requirement Target
DSP bypass Bit-identical where possible
Internal processing Float32 minimum
Offline analysis Float64 recommended
Output clipping 0 clipped samples
True-peak ceiling -1.0 dBTP for lossy output paths
Loudness measurement ITU-R BS.1770 / EBU R128 compatible
Sample-rate conversion One conversion maximum per playback path
Real-time callback No locks, allocation, file I/O, network I/O, or logging
User control Pure Mode and Enhanced Mode
Regression tests Null test, sweep test, clipping test, loudness test

ITU-R BS.1770-5 defines algorithms for programme loudness and true-peak audio measurement.

AES loudness resources list ITU-R BS.1770-5 and AES streaming loudness recommendations.


4. Recommended Apple-native plus C++23 architecture

Input file
→ Apple decoder or FFmpeg fallback
→ Float32 PCM buffer
→ optional sample-rate conversion
→ pre-gain
→ parametric EQ
→ loudness gain
→ true-peak-aware limiter
→ Core Audio output

For analysis:

Input file
→ decode to analysis buffer
→ loudness scan
→ peak scan
→ true-peak scan
→ waveform cache
→ spectrum cache
→ metadata database

Do all analysis outside the real-time render callback.


5. Apple-native stack versus open-source stack

Layer Apple-native choice Open-source choice Recommendation
Audio output Core Audio / AVAudioEngine JUCE, miniaudio, PortAudio Use Apple native.
Audio graph AVAudioEngine JUCE graph Use Apple native for macOS-only apps.
Custom effect Audio Unit version 3 JUCE plug-in wrapper Use AUv3 if Apple-only.
Real-time DSP C++23 kernel JUCE DSP, custom C++ Use custom C++23.
Vector DSP Accelerate / vDSP Eigen, FFTW, custom SIMD Use vDSP first.
Loudness scan Custom code libebur128 Use libebur128.
Resampling AVAudioConverter soxr, libsamplerate, SSRC Native first. Use soxr or SSRC for expert offline mode.
Codec expansion AVFoundation FFmpeg Add FFmpeg only when needed.
AI inference Core ML / MLX PyTorch, ONNX Runtime Use Core ML for shipping. Use MLX for research.

6. Hardware acceleration on Apple Silicon

Apple Silicon feature Use for music rendering Practical recommendation
CPU performance cores Real-time DSP Use for EQ, limiter, gain, routing, crossfeed.
CPU efficiency cores Background work Use for scanning, metadata, waveform cache.
Accelerate / vDSP Filters, FFT, convolution Use for math-heavy DSP before writing custom SIMD.
Metal GPU Visualizers, long convolution, batch analysis Use outside the real-time callback first.
Neural Engine Core ML inference Use for offline or buffered AI features.
Unified memory Shared CPU/GPU/ML data Useful for large waveforms and ML models.
M5 GPU Neural Accelerators GPU-backed AI paths Useful for AI, not needed for normal playback.
Media engine Video codecs Not relevant to normal music playback.

Apple’s M1 MacBook Air has an 8-core CPU, 7-core or 8-core GPU, and 16-core Neural Engine.

Apple says the M1 Neural Engine supports 11 trillion operations per second.

Apple announced M5 with a 10-core CPU, GPU Neural Accelerators, 16-core Neural Engine, and 153 GB/s unified memory bandwidth.

For EQ, limiter, and loudness gain, M1 is enough.

For AI stem separation and large local models, M5 gives more headroom.


7. DSP features that improve rendering quality

Feature Use it? Why it matters Implementation tip
Pure Mode Yes Gives reference playback. Bypass all DSP.
Float32 pipeline Yes Prevents internal clipping during gain and EQ. Keep the live path in floating point.
Float64 analysis Yes Improves offline measurement precision. Use for loudness and test tools.
Pre-gain Yes Creates headroom before EQ boosts. Reduce gain before positive bands.
Parametric EQ Yes More controlled than fixed-band EQ. Use stable biquad filters.
Limiter Yes Prevents clipping after EQ and gain. Put it last.
True-peak meter Yes Finds inter-sample peaks. Oversample the meter path.
Loudness normalization Yes Reduces volume jumps between tracks. Use ITU-R BS.1770 / EBU R128.
Dither Export only Reduces quantization distortion. Use only when reducing bit depth.
Oversampling Selective Reduces nonlinear artifacts. Use around saturation or clipping only.
Room correction Optional Corrects speaker-room response. Require measurements. Avoid blind boosts.
Crossfeed Optional Helps headphone listening. Make it headphone-only.
Stereo widening Risky Can damage mono compatibility. Keep bass mostly mono below 120 Hz.
AI enhancement Optional Can restore or invent content. Make it offline and reversible.

8. Resampling and format conversion

Resampling affects quality when input sample rate differs from output device rate.

Path Recommendation
Live playback Prefer device-native rate or one Apple conversion.
Offline export Use highest-quality converter settings.
Audiophile mode Avoid unnecessary conversion.
Crossfade between rates Convert once into the session rate.
AI model input Convert a copy, not the live stream.

Use AVAudioConverter first for live playback.

Use soxr or SSRC when you need more offline control.

soxr exposes phase response, preserved bandwidth, aliasing, and rejection parameters.

SSRC describes itself as an audiophile-grade sample-rate converter for PCM WAV files.

Treat project descriptions as claims until you test them in your own harness.


9. Loudness normalization and true peak

Use ITU-R BS.1770-compatible measurement.

Do not normalize by sample peak only.

Peak normalization does not track perceived loudness.

Recommended metadata:

Field Purpose
Integrated loudness Overall perceived loudness.
Loudness range Dynamic range estimate.
Sample peak Highest sample value.
True peak Reconstructed peak estimate.
Track gain Gain for shuffled playback.
Album gain Gain preserving album balance.

Playback rule:

decoded audio
→ track or album gain
→ limiter
→ output

Use album gain for albums.

Use track gain for shuffled playlists.

Target a true-peak ceiling around -1.0 dBTP when the output path includes lossy encoding or device conversion.


10. Equalizer design

For version 1, use a parametric EQ.

A parametric EQ gives frequency, gain, and Q control.

Q means filter sharpness.

Recommended EQ path:

pre-gain
→ high-pass filter if needed
→ low-shelf
→ parametric bands
→ high-shelf
→ limiter

Implementation tips:

Issue Fix
Clicks during slider movement Smooth parameters over 5–50 ms.
Clipping after boosts Add pre-gain.
Filter instability Clamp frequency, gain, and Q.
Poor stereo image Apply matched filters to left and right.
CPU spikes Precompute coefficients outside callback.

Use vDSP or a proven C++ biquad implementation.

Add sweep tests and impulse tests.


11. Limiter design

The limiter protects output quality after EQ and loudness gain.

Recommended first version:

Parameter Value
Ceiling -1.0 dBFS or -1.0 dBTP
Lookahead 1–5 ms
Release 50–200 ms
Oversampling 2x or 4x for true-peak mode
Position Last DSP stage

Avoid aggressive compression in default mode.

Add a separate “Gym Mode” for noisy rooms.

Gym Mode can use compression because fidelity is not the main goal.


12. Room correction

Room correction can improve speaker playback more than changing the player engine.

It also carries higher risk than EQ.

Technique Benefit Risk
Minimum-phase EQ Low latency Does not fix all time-domain problems.
Linear-phase FIR Corrects phase and magnitude Adds latency and pre-ringing.
Mixed-phase correction Balances both More complex design.
Partitioned convolution Runs long filters efficiently More complex buffering.
Multi-position correction Better listening area Needs multiple measurements.

Practical rule:

measurement import
→ target curve
→ correction filter
→ limiter
→ output

Do not boost deep nulls.

Deep nulls are often caused by cancellation.

Boosting them wastes headroom and can stress speakers.

Use cuts before boosts.


13. AI-based audio enhancement

AI can improve poor recordings.

AI can also invent content that was not in the source.

That conflicts with audiophile transparency.

AI approach Usefulness Production recommendation
Music source separation Remix, karaoke, analysis Offline feature.
Audio super-resolution Low-bandwidth sources Offline preview only.
AI denoise User recordings Not default for mastered music.
Differentiable DSP Controllable neural effects Research path.
Neural audio codecs Compression research Not v1 playback.
Audio tagging Library intelligence Good background feature.

Open-Unmix is a peer-reviewed open-source reference implementation.

Hybrid Transformer Demucs reports stronger separation results than older Demucs variants.

AudioSR is a diffusion-based audio super-resolution model.

DDSP combines classical DSP blocks with neural networks.

For your product, put these under “AI Lab Mode.”

Do not put them in default playback.


14. Apple-specific AI deployment

Option Best use Notes
Core ML Shipping models inside macOS app Uses CPU, GPU, and Neural Engine.
coremltools Convert PyTorch/TensorFlow models Use during build or research flow.
MLX Apple Silicon research and prototypes Supports Python, C++, C, and Swift APIs.
PyTorch MPS Training and experiments Good for research, not final app runtime.
Metal Performance Shaders Graph Custom graph compute Use when Core ML is not enough.

Use Core ML when you ship.

Use MLX when you experiment locally.

Do not run Core ML inference inside the real-time callback.


15. Real-time safety rules

Rule Reason
No heap allocation Allocation can block.
No locks Lock contention causes dropouts.
No file I/O Disk latency is not bounded.
No network I/O Network latency is not bounded.
No logging Logging can lock or allocate.
No Core ML inside callback Inference latency is not guaranteed.
No Swift object churn Reference counting adds unpredictable work.
Use preallocated buffers Keeps render time bounded.
Use lock-free parameter exchange Keeps UI updates safe.

RealtimeWatchdog checks for unsafe Core Audio thread activity.

It flags locks, memory allocation, Objective-C use, file I/O, and network I/O.

Use it as a development guardrail.


16. Testing plan

Test Method Pass condition
Bypass null test Output minus input Near-zero residual where no conversion occurs.
Frequency response Sweep or impulse Matches target curve.
EQ stability Parameter sweep No clicks or unstable filters.
Clipping test Full-scale sine and music 0 clipped samples.
True-peak test Oversampled meter Peak below ceiling.
Loudness test ITU-R BS.1770 vectors Matches reference library.
Resampler test Sweep conversion No obvious aliasing or passband ripple.
CPU test 128-frame buffer No underruns under UI load.
Latency test Loopback measurement Document measured latency by device.
AI test A/B comparison AI must be optional and reversible.

Use objective tests first.

Then run listening tests.

Do not rely on listening alone.


17. Recommended product modes

Mode Processing User promise
Pure Mode Decode → output We do not change the sound.
Enhanced Mode EQ → loudness gain → limiter We improve consistency and prevent clipping.
Room Mode Room correction → limiter We correct speaker-room response from measurements.
Headphone Mode Crossfeed → EQ → limiter We reduce hard stereo separation.
AI Lab Mode Offline AI tools Experimental restoration and separation.

For version 1, build:

Pure Mode
Enhanced Mode
Analysis Mode

Add Room Mode after measurement import works.

Add AI Lab Mode after you build A/B comparison and rollback.


18. Implementation roadmap

Phase Build
1 SwiftUI app, playlist, playback, AVAudioEngine output
2 C++23 DSP kernel with bypass, gain, limiter
3 Parametric EQ using biquad filters
4 Offline loudness scan using libebur128
5 Waveform and spectrum cache using vDSP FFT
6 True-peak scan and clipping reports
7 Optional soxr or SSRC offline resampling
8 Room correction import using FIR filters
9 Core ML or MLX offline AI experiments
10 AUv3 plug-in version for Logic Pro or GarageBand support

19. Recommended v1 dependency list

Component Choice Reason
UI SwiftUI Native macOS UI.
Audio graph AVAudioEngine Native graph and output path.
Output Core Audio through Apple APIs Best Apple platform integration.
DSP core C++23 Predictable real-time code.
Math acceleration Accelerate / vDSP Apple-optimized vector DSP.
Loudness libebur128 Implements EBU R128 style measurement.
Codec fallback FFmpeg Use only when Apple APIs miss a format.
Resampling expert mode soxr or SSRC Better control for offline conversion.
AI shipping Core ML Best Apple app deployment path.
AI research MLX Apple Silicon research framework.

20. Actively maintained C/C++ audio libraries with release and community signals

This section was added on 2026-06-17.

Maintenance was judged using recent releases, visible repository activity, package-manager updates, security response, and project adoption.

Community signal was judged using GitHub visibility, ecosystem use, packaging availability, and use by other audio projects.

20.1 Strong maintenance signals

Library Category Maintenance signal Community signal Recommendation for your player
JUCE C++ audio app and plug-in framework GitHub releases show JUCE 8.x releases, including 8.0.12 in the current release feed. Used across commercial and open-source audio plug-ins. Supports VST3, AU, AUv3, AAX, and LV2. Use only if cross-platform release or plug-in hosting is a product goal.
miniaudio C/C++ playback, capture, decoding, and mixing Latest release page showed v0.11.25 on 2026-03-04. Single-file design, no external dependencies, and broad desktop/mobile platform support. Good fallback for tools or cross-platform experiments. For Apple-only output, use AVAudioEngine/Core Audio.
FFmpeg Codec, container, conversion, and filter stack Official FFmpeg page announced 8.1 “Hoare” on 2026-03-16. Download page also showed maintained release branches. Industry-standard multimedia stack with broad codec/container support. Use as a fallback for formats not covered by Apple APIs. Check LGPL/GPL build configuration.
Faust DSP language and C++ code generation GitHub release page showed recent 2.8x releases. Faust-to-CLAP work appeared in current research and tooling. Used for DSP prototyping, code generation, and audio research. Use to prototype EQ, filters, and effects, then port or generate C++ for the DSP kernel.
Rubber Band Library Time-stretching and pitch-shifting Official page announced v4.0 on 2024-10-25. Used by audio/video tools for tempo and pitch changes. Use for optional tempo and pitch features. Keep it out of default hi-fi playback.
KissFFT Fast Fourier Transform GitHub and distribution metadata show 131.x packaging, including 131.2.0 in Fedora/Arch metadata. Small, portable, BSD-licensed FFT library. Use as a portable FFT fallback. On Apple Silicon, prefer Accelerate/vDSP first.
PFFFT FFT and fast convolution Repository and package metadata show active packaging updates in 2026. Small FFT library with ARM NEON support, including Apple Silicon. Good portable FFT/convolution fallback if you do not want to depend only on vDSP.
r8brain-free-src Sample-rate conversion Repository and documentation were current enough to be useful; project had 2026 issue activity. Known in audio-development communities for high-quality resampling. MIT license helps commercial apps. Use for offline or expert-mode sample-rate conversion. For live playback, start with AVAudioConverter.
dr_libs WAV, FLAC, MP3 single-file decoders vcpkg listed 2026 package versions. Security issues were discussed publicly in 2026, with fix guidance to update to latest master. Common in games, tools, and lightweight C/C++ projects. Use for test tools or simple import paths. Avoid old versions for untrusted files.
Essentia C++ audio and music information retrieval PyPI showed a 2026 release for the Python package backed by the C++ library. Used in MIR research and music-analysis tooling. MIR means music information retrieval. Useful for offline analysis. AGPLv3 license can be incompatible with closed-source commercial apps.

20.2 Useful but not “actively maintained” by release cadence

Library Category Maintenance signal Why still useful Recommendation
libebur128 Loudness and true-peak measurement Latest GitHub release was older, around 2021. EBU R128 and ITU-R BS.1770 are stable standards. Use with regression tests against known loudness vectors.
libsamplerate Sample-rate conversion Latest GitHub release shown was 0.2.2 from 2021. Mature library with BSD-style licensing. Fine for tooling. Prefer r8brain, soxr, or AVAudioConverter for production decisions.
soxr Sample-rate conversion Upstream release cadence is old, but it remains packaged and downloaded. Exposes phase, bandwidth, aliasing, and rejection controls. Good for offline expert conversion after you test it.
libsndfile WAV, AIFF, and sampled audio file I/O Latest GitHub release shown was 1.2.2 from 2023-08-13. Mature and widely packaged. Use for test tools and offline exports.
RtAudio Cross-platform real-time audio I/O Official latest release was 6.0.1 from 2023-08-01. Stable cross-platform API for ALSA, JACK, PulseAudio, OSS, macOS, and Windows APIs. Use only if you need cross-platform audio I/O. Do not use for Apple-only output.
zita-convolver Convolution Older cadence and GPL license. Useful technical design for partitioned convolution. Use as reference or in GPL-compatible products.

20.3 My current shortlist

For your Apple Silicon C++23 + Swift player:

Need Pick Why
macOS output AVAudioEngine / Core Audio Native device integration and routing.
Real-time DSP Your C++23 engine Keeps the render path predictable.
Vector math / FFT Accelerate / vDSP Apple-optimized path for Apple Silicon.
Portable FFT fallback PFFFT or KissFFT Small and portable.
EQ/filter prototyping Faust Fast DSP iteration and C++ generation.
Loudness scan libebur128 Implements the standard workflow.
Offline resampling r8brain-free-src MIT-licensed C++ resampler.
Codec fallback FFmpeg Broadest codec/container support.
Simple test decoders dr_libs Single-file WAV/FLAC/MP3 tooling.
Time-stretch/pitch Rubber Band Specialist library for tempo and pitch.
Offline music analysis Essentia Strong analysis coverage, but check AGPLv3.
Cross-platform app future JUCE Full app and plug-in ecosystem.

20.4 Decision

Do not replace Apple’s audio output path with a cross-platform library unless cross-platform release is required.

Use this stack first:

SwiftUI
→ AVAudioEngine / Core Audio
→ C++23 DSP kernel
→ Accelerate / vDSP
→ libebur128
→ r8brain-free-src
→ FFmpeg fallback only when needed

Use JUCE later only if you decide to ship Windows/Linux builds or plug-in formats.


21. Final recommendation

For an Apple-only music player:

Apple native output
+ C++23 real-time DSP
+ Accelerate / vDSP
+ selected open-source libraries
+ optional offline AI

This gives the best balance between fidelity, control, latency, power efficiency, and maintenance.

Do not enhance every track by default.

Make every processing step measurable, visible, and reversible.


Verified references

The links below were checked against official pages, standards pages, paper pages, or original repositories on 2026-06-17.

Apple official documentation and platform material

  1. Apple AVAudioEngine documentation
    https://developer.apple.com/documentation/avfaudio/avaudioengine

  2. Apple AVAudioConverter documentation
    https://developer.apple.com/documentation/avfaudio/avaudioconverter

  3. Apple Technical Note TN3136: AVAudioConverter sample-rate conversion
    https://developer.apple.com/documentation/technotes/tn3136-avaudioconverter-performing-sample-rate-conversions

  4. Apple Core ML documentation
    https://developer.apple.com/documentation/coreml

  5. Apple Core ML overview
    https://developer.apple.com/machine-learning/core-ml/

  6. Apple Accelerate framework
    https://developer.apple.com/documentation/accelerate

  7. Apple vDSP audio unit sample
    https://developer.apple.com/documentation/accelerate/creating-an-audio-unit-extension-using-the-vdsp-library

  8. Apple FFT documentation
    https://developer.apple.com/documentation/accelerate/fast-fourier-transforms

  9. Apple Metal overview
    https://developer.apple.com/metal/

  10. Apple Metal Performance Shaders
    https://developer.apple.com/documentation/metalperformanceshaders

  11. Apple Metal Performance Shaders Graph
    https://developer.apple.com/documentation/metalperformanceshadersgraph

  12. Apple PyTorch MPS backend page
    https://developer.apple.com/metal/pytorch/

  13. Apple AUv3 custom audio effects sample
    https://developer.apple.com/documentation/avfaudio/creating-custom-audio-effects

  14. Apple signal generator sample
    https://developer.apple.com/documentation/avfaudio/building-a-signal-generator

  15. Swift C++ interoperability
    https://www.swift.org/documentation/cxx-interop/

  16. Apple M1 announcement
    https://www.apple.com/newsroom/2020/11/apple-unleashes-m1/

  17. Apple M1 MacBook Air technical specification
    https://support.apple.com/en-lk/111883

  18. Apple M5 announcement
    https://www.apple.com/newsroom/2025/10/apple-unleashes-m5-the-next-big-leap-in-ai-performance-for-apple-silicon/

  19. Apple M5 MacBook Air technical specification
    https://support.apple.com/en-lk/126320

Standards and engineering references

  1. ITU-R BS.1770-5: loudness and true-peak measurement
    https://www.itu.int/rec/R-REC-BS.1770

  2. ITU-R BS.1770-5 PDF
    https://www.itu.int/dms_pubrec/itu-r/rec/bs/R-REC-BS.1770-5-202311-I!!PDF-E.pdf

  3. AES loudness references and standards list
    https://aes.org/resources/audio-topics/loudness-project/resources-and-references/

  4. EBU R128 publication page
    https://tech.ebu.ch/publications/r128/

Open-source DSP and audio libraries

  1. libebur128
    https://github.com/jiixyj/libebur128

  2. soxr / SoX Resampler library
    https://sourceforge.net/projects/soxr/

  3. libsamplerate / Secret Rabbit Code
    https://github.com/libsndfile/libsamplerate

  4. Shibatch Sample Rate Converter
    https://github.com/shibatch/ssrc

  5. FFmpeg codecs documentation
    https://ffmpeg.org/ffmpeg-codecs.html

  6. FFmpeg legal and license information
    https://www.ffmpeg.org/legal.html

  7. JUCE framework
    https://github.com/juce-framework/JUCE

  8. AudioKit
    https://github.com/AudioKit/AudioKit

  9. RealtimeWatchdog
    https://github.com/TheAmazingAudioEngine/RealtimeWatchdog

Apple Silicon machine learning projects

  1. MLX official repository
    https://github.com/ml-explore/mlx

  2. MLX official project page
    https://opensource.apple.com/projects/mlx

  3. MLX unified memory documentation
    https://ml-explore.github.io/mlx/build/html/usage/unified_memory.html

  4. coremltools
    https://github.com/apple/coremltools

  5. coremltools PyTorch conversion workflow
    https://apple.github.io/coremltools/docs-guides/source/convert-pytorch-workflow.html

  6. PyTorch MPS backend documentation
    https://docs.pytorch.org/docs/stable/notes/mps.html

Peer-reviewed or research material

  1. Open-Unmix JOSS paper
    https://joss.theoj.org/papers/10.21105/joss.01667

  2. Open-Unmix PyTorch repository
    https://github.com/sigsep/open-unmix-pytorch

  3. Open-Unmix project page
    https://sigsep.github.io/open-unmix/

  4. Hybrid Transformer Demucs paper
    https://arxiv.org/abs/2211.08553

  5. Demucs repository
    https://github.com/facebookresearch/demucs

  6. AudioSR paper
    https://arxiv.org/abs/2309.07314

  7. AudioSR project page
    https://audioldm.github.io/audiosr/

  8. AudioSR repository
    https://github.com/haoheliu/versatile_audio_super_resolution

  9. DDSP OpenReview paper
    https://openreview.net/forum?id=B1x1ma4tDr

  10. DDSP repository
    https://github.com/magenta/ddsp

  11. Review of differentiable DSP for music and speech synthesis
    https://doi.org/10.3389/frsip.2023.1284100

  12. Room Response Equalization — A Review
    https://www.mdpi.com/2076-3417/8/1/16

  13. Real-time low-latency music source separation using Hybrid Spectrogram-TasNet
    https://arxiv.org/abs/2402.17701

  14. PEAQ reassessment paper
    https://arxiv.org/abs/2212.01467

Additional references for actively maintained C/C++ libraries

  1. JUCE GitHub releases
    https://github.com/juce-framework/JUCE/releases

  2. JUCE project repository
    https://github.com/juce-framework/JUCE

  3. miniaudio GitHub releases
    https://github.com/mackron/miniaudio/releases

  4. miniaudio official project page
    https://miniaud.io/

  5. FFmpeg official website
    https://www.ffmpeg.org/

  6. FFmpeg official download page
    https://ffmpeg.org/download.html

  7. Faust GitHub releases
    https://github.com/grame-cncm/faust/releases

  8. Faust project repository
    https://github.com/grame-cncm/faust

  9. Rubber Band Library official page
    https://breakfastquay.com/rubberband/

  10. Rubber Band GitHub mirror
    https://github.com/breakfastquay/rubberband

  11. KissFFT repository
    https://github.com/mborgerding/kissfft

  12. KissFFT Fedora package metadata
    https://rpmfind.net/linux/RPM/fedora/44/s390x/k/kiss-fft-131.2.0-1.fc44.s390x.html

  13. PFFFT repository
    https://github.com/marton78/pffft

  14. PFFFT FreeBSD package metadata
    https://www.freshports.org/math/pffft

  15. r8brain-free-src repository
    https://github.com/avaneev/r8brain-free-src

  16. r8brain-free-src documentation
    https://www.voxengo.com/public/r8brain-free-src/Documentation/

  17. dr_libs repository
    https://github.com/mackron/dr_libs

  18. drlibs vcpkg package metadata
    https://vcpkg.io/en/package/drlibs.html

  19. dr_libs 2026 security advisory reference
    https://github.com/marlinkcyber/advisories/blob/main/advisories/MCSAID-2026-001-dr-libs-heap-overflow.md

  20. RtAudio official page
    https://caml.music.mcgill.ca/~gary/rtaudio/

  21. Essentia repository
    https://github.com/MTG/essentia

  22. Essentia PyPI release page
    https://pypi.org/project/essentia/

  23. libsndfile GitHub releases
    https://github.com/libsndfile/libsndfile/releases/

Verification notes

  • Apple sources were checked against Apple Developer, Apple Support, or Apple Newsroom pages.
  • ITU and AES loudness references were checked against ITU and AES pages.
  • Open-source projects were checked against original GitHub or SourceForge repositories.
  • Peer-reviewed entries were checked against JOSS, Frontiers, MDPI, or OpenReview pages.
  • arXiv papers were treated as research preprints unless a reviewed venue was also linked.
  • Repository claims were not treated as proof of audio quality without independent testing.