Modern Adaptive Differential Pulse-Code Modulation

ADPCM didn't get worse. Everyone just stopped trying.
A small, fast, waveform audio codec that fixes what IMA ADPCM got wrong.
No transforms. No psychoacoustics. No entropy coding. No excuses.

The obvious replacement for IMA ADPCM when you control both ends.

what is it?

The math of FLAC, but the constraints of ADPCM.
~500 lines of pure C.

under the hood

MADPCM implements a fixed-point 4th-order Linear Predictive Coder (LPC) derived via Levinson-Durbin recursion.
A pre-encoding simulation and heuristics loop validates coefficients using Zero Crossing Rate (ZCR) and signal-to-residual energy ratios, triggering fallback mechanisms on model instability.
Quantization utilizes a Viterbi trellis search ($K=4$) for global optimization or a greedy O(1) inverse lookahead for speed.
Stereo coupling employs adaptive Mid-Side coding, dynamically switching based on inter-channel energy correlation.
The core uses Gaussian-optimized step tables and is strictly integer-based to support toasters (read: freestanding environments).

The decoder runs entirely in registers, no stack spilling provided it's compiled with optimizations (tested with mingw gcc 15.1.0).

warning

This is work in progress, expect breaking changes every other commit.

the numbers

Measured on 16-bit PCM sources:

IMA ADPCM (ffmpeg):

File	RMSE	MAE	PSNR	SNR	Enc Time	Dec Time
Marche_Persanne.wav	49.81	3345	56.36 dB	32.38 dB	0.119s	0.058s
Egyptischer_Marsch.wav	56.51	5872	55.27 dB	32.70 dB	0.256s	0.145s
suppe_poet_and_peasant.wav	65.80	8322	53.94 dB	31.26 dB	0.636s	0.348s
the_four_seasons.wav	47.71	2630	56.74 dB	36.19 dB	3.927s	2.150s

ADPCM-XQ (lookahead=8):

File	RMSE	MAE	PSNR	SNR	Enc Time
Marche_Persanne.wav	42.75	1228	57.69 dB	33.71 dB	17.486s
Egyptischer_Marsch.wav	47.31	1851	56.81 dB	34.25 dB	40.850s
suppe_poet_and_peasant.wav	54.54	1640	55.58 dB	32.89 dB	102.432s
the_four_seasons.wav	44.67	778	57.31 dB	36.76 dB	607.841s

MADPCM (slow):

File	RMSE	MAE	PSNR	SNR	Enc Time	Dec Time
Marche_Persanne.wav	27.30	343	61.59 dB	37.61 dB	0.178s	0.008s
Egyptischer_Marsch.wav	32.09	623	60.18 dB	37.62 dB	0.384s	0.016s
suppe_poet_and_peasant.wav	37.05	810	58.93 dB	36.25 dB	0.991s	0.038s
the_four_seasons.wav	25.00	222	62.35 dB	41.80 dB	6.065s	0.214s

MADPCM (fast; no lookahead):

File	RMSE	MAE	PSNR	SNR	Enc Time
Marche_Persanne.wav	29.01	512	61.06 dB	37.08 dB	0.020s
Egyptischer_Marsch.wav	34.01	600	59.68 dB	37.11 dB	0.054s
suppe_poet_and_peasant.wav	39.36	802	58.41 dB	35.72 dB	0.115s
the_four_seasons.wav	26.37	271	61.89 dB	41.34 dB	0.680s

All tested on a Ryzen 5 3600XT running Windows Server 2025 Datacenter build 26100.32230
Decode speed for MADPCM is the same on slow and fast, this affects only the encoder.
Encode speed for ADPCM-XQ includes file I/O, tested on a ramdisk to minimize potential penalty.
File size reduction: ~74% vs WAV, 4.48% overhead compared to IMA ADPCM.

test file details

All downloaded from youtube with premium at max quality, converted to wav.
Egyptischer_Marsch.wav 43,5 MB Length: 03:57 Sample rate: 48 kHz Sample size: 16 bit Stereo
Marche_Persanne.wav 18,7 MB Length: 01:51 Sample rate: 48 kHz Sample size: 16 bit Stereo
suppe_poet_and_peasant.wav 107 MB Length: 09:49 Sample rate: 48 kHz Sample size: 16 bit Stereo
the_four_seasons.wav 665 MB Length: 01:00:33 Sample rate: 48 kHz Sample size: 16 bit Stereo

EBU Sound Quality Assessment Material

MADPCM (slow):

Audio Sample	RMSE	MAE	PSNR	SNR
Trumpet	16.0853	259	66.18	41.73
Castanets	80.8316	2502	52.16	20.94
Side Drum	16.3826	317	66.02	39.33
Glockenspiel	45.8915	1782	57.07	30.78
Vibraphone	17.9936	256	65.21	42.56
Grand Piano	15.4270	213	66.54	43.23
Harpsichord	30.6090	770	60.59	27.78
Female Speech	45.9353	1135	57.07	33.53

IMA ADPCM (ffmpeg):

Audio Sample	RMSE	MAE	PSNR	SNR
Trumpet	34.8689	2434	59.46	35.01
Castanets	109.273	15673	49.54	18.33
Side Drum	24.5968	2407	62.49	35.80
Glockenspiel	83.6360	20459	51.86	25.57
Vibraphone	23.4344	865	62.91	40.26
Grand Piano	16.6159	1233	65.90	42.58
Harpsichord	39.5648	3725	58.36	25.55
Female Speech	72.1851	3461	53.14	29.61

how other underpowered hardware behaves

samsung s21:

~/madpcm $ ./test.out encode slow suppe_poet_and_peasant.wav suppe_poet_and_peasant_ad.wav
Loaded suppe_poet_and_peasant.wav: 48000 Hz, Stereo
Encoded in 4.128757 seconds (slow mode)
~/madpcm $ ./test.out encode fast suppe_poet_and_peasant.wav suppe_poet_and_peasant_ad.wav
Loaded suppe_poet_and_peasant.wav: 48000 Hz, Stereo
Encoded in 0.311960 seconds (fast mode)
~/madpcm $ ./test.out decode suppe_poet_and_peasant_ad.wav suppe_poet_and_peasant_d.wav
Loaded suppe_poet_and_peasant_ad.wav: 2 channels, 48000 Hz, 28272640 samples
Decoded in 0.079933 seconds

raspberry pi 4:

maksw@raspberrypi:~/madpcm $ ./test.out encode slow suppe_poet_and_peasant.wav suppe_poet_and_peasant_ad.wav
Loaded suppe_poet_and_peasant.wav: 48000 Hz, Stereo
Encoded in 4.307982 seconds (slow mode)
maksw@raspberrypi:~/madpcm $ ./test.out encode fast suppe_poet_and_peasant.wav suppe_poet_and_peasant_ad.wav
Loaded suppe_poet_and_peasant.wav: 48000 Hz, Stereo
Encoded in 0.430510 seconds (fast mode)
maksw@raspberrypi:~/madpcm $ ./test.out decode suppe_poet_and_peasant_ad.wav suppe_poet_and_peasant_d.wav
Loaded suppe_poet_and_peasant_ad.wav: 2 channels, 48000 Hz, 28272640 samples
Decoded in 0.213578 seconds

how a 486 behaves (dosbox-x)

Environment: Emulated 486 (25,000 cycles), 15s Stereo WAV @ 44.1kHz.

C:\>mad.exe encode fast 37.wav 37_ad.wav
Loaded 37.wav: 44100 Hz, Stereo
Encoded in 10.659341 seconds (fast mode)

C:\>mad.exe decode 37_ad.wav 37_d.wav
Loaded 37_ad.wav: 2 channels, 44100 Hz, 661504 samples
Decoded in 2.747253 seconds

Encoder runs faster than realtime.
Decoder is ~5.4× realtime.
scary shit

does it sound better?

To me. I think. Better than IMA ADPCM.
No formal ABX tests were done yet.

In my personal testing (Sony WF-1000XM4 connected via LDAC) compared to the source, MADPCM adds a very small, mostly uncorrelated noise floor. Compared to IMA ADPCM, it sounds much cleaner.

how 2 use it

look in Reference.md

what this does not care about

• standards
• legacy compatibility
• your favorite codec comparison

how to contribute

Test it. Report bugs. Open PRs.
4 spaces, K&R styling.

q&a

Q: Why make this?
A: IMA ADPCM works. This works better. IMA's everywhere because it's cheap and simple, not because it's good.

Q: Is this better than Opus/AAC/MP3?
A: If you squint hard enough and just look at the numbers. This is a scalpel, not a swiss army knife.

Q: Is this a competitor to Opus?
A: No.

Q: What's the use case?
A: Smaller-than-WAV files, predictable latency, trivial decoding cost.

Q: Are there transforms or psychoacoustic models?
A: No. Conceptually inspired by lossless predictive codecs like FLAC.

Q: Why is encoding slow?
A: Quality costs time.

Q: Will you add [feature]?
A: Depends. Open an issue.

Q: Can I use this in production?
A: No warranty. Assume breaking changes until it's finished.
If you really want to use it, at least keep track of the commit you used.