The NDS contains a built-in ADPCM decoder. However discussing the encoder architecture with Psy_Commando and reviewing notes on the subject created by martin korth here http://problemkaputt.de/gbatek-ds-sound-notes.htm, revealed that there appears to be a slight discrepancy between the standard IMA suggested implementation and the actual implementation: this fork of ADPCM-XQ seeks to solve that with an encoder dedicated to NDS audio.
During the step where the difference is added to the running prediction, there is the possibility of the final sample value falling outside of the range of a 16-bit signed integer, and so the NDS, like the reference implementation, handles this by clipping the final sample value to a good range. However this range is different between the IMA reference and the NDS implementation. While the reference clips the value to [-32768, 32767], utilizing the entire range of the 16-bit signed int, the NDS decoder instead clips to [-32767, 32767], preferring the symmetry. ADPCM is cumulative, so this can lead to unexpected results. The main change here is to modify this clipping to match that of the NDS.
Below is the original README for ADPCM-XQ:
Xtreme Quality ADPCM Encoder/Decoder
Copyright (c) 2024 David Bryant.
All Rights Reserved.
Distributed under the BSD Software License.
While very popular at the end of the last century, ADPCM is no longer a common audio encoding format, and is certainly not recommended as a general purpose encoder. However, it requires minimal CPU resources for decoding, and so still is ideally suited for certain embedded games and applications that contain canned audio samples.
This encoder combines two different techniques to achieve higher quality than existing ADPCM encoders while remaining fully compatible with standard decoders. The first is dynamic noise shaping, which shifts the quantization noise up or down in frequency based on the spectrum of the source signal. This technique is identical to the algorithm used in WavPack's lossy mode and can make any audible quantization noise much less annoying (or, in some cases, inaudible).
The other technique is "lookahead" in which the encoder exhaustively searches ahead to find the optimum coding sequence based on future samples. This process can reduce the quantization noise from about 1 to 10 dB (depending on the source) and also reduces or eliminates the harmonic content in the noise that sometimes plagues ADPCM. Unfortunately, at its maximum settings this can be very slow, but this should be relatively irrelevant if the encoder is being used to generate canned samples.
Adpcm-xq consists of three standard C files and a header file (adpcm-lib.h). It can be used as a stand-alone command-line program implemented in adpcm-xq.c, or the library, which consists of adpcm-lib.c and adpcm-dns.c, can be built into and utilized by another application. The library portion has been designed with maximum portability in mind and should work correctly even on 16-bit and big-endian architectures.
The latest version of adpcm-xq has many enhancements including greatly improved performance of the conversion and the ability to calculate and display the quantization noise introduced in the operation. See all the details here.
There are several forms and variations of IMA ADPCM encoding. The one handled by adpcm-xq is the canonical one used in Microsoft WAV files. The audio is divided into fixed-sized blocks that include a 4-byte header (or 8-byte for stereo) that includes the first sample. The size of the blocks is stored in the WAV header and the audio nibbles are ordered least-significant temporally first.
The latest version of adpcm-xq also includes 2-bit, 3-bit and 5-bit ADPCM. These are not well-supported at all (and some support is buggy) but they might be useful if this library is decoding them in situations where more compression or higher quality is desired. BTW, Rockbox provides excellent support for them!
Some applications like games and consoles that decode IMA ADPCM in hardware or microcode use modified versions of IMA ADPCM that do not use headers and do not divide the audio into blocks or frames of any kind. The decoding parameters are simply initialized to zero and the audio nibbles continue uninterrupted to the end of the clip. I have created an experimental version that will generate two variations of this data. One is standard nibble order and the other is reversed nibble order (sometimes called Intel/DVI4 or ADP4 and is the format used in AIFF files). These formats are only writable as "raw" by the adpcm-xq command-line program because they are not representable in WAV files, and they cannot be decoded by the program either (for the same reason), however the library itself handles them. The experimental branch is here.
To build the command-line tool (ADPCM-XQ) on Linux:
$ gcc -O3 *.c -lm -o adpcm-xq
on Darwin/Mac:
$ cmake -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" . ; make
on MS Visual Studio:
C:\cl -O3 adpcm-xq.c adpcm-lib.c adpcm-dns.c
ADPCM-XQ Xtreme Quality IMA-ADPCM WAV Encoder / Decoder Version 0.5
Copyright (c) 2024 David Bryant. All Rights Reserved.
Usage: ADPCM-XQ [-options] infile.wav outfile.wav
Operation: conversion is performed based on the type of the infile
(either encode 16-bit PCM to 4-bit IMA-ADPCM or decode back)
Options: -[0-16]= encode lookahead samples (default = 3, max = 16)
-b<n> = override auto block size, 2^n bytes (n = 8-15)
-d = decode only (fail on WAV file already PCM)
-e = encode only (fail on WAV file already ADPCM)
-f = encode flat noise (no noise shaping, aka -s0.0)
-h = display this help message
-n = measure and report quantization noise
-q = quiet mode (display errors only)
-r = raw output (little-endian, no WAV header written)
-s<n> = override default noise shaping, (-1.0 < n <= 1.0)
-v = verbose (display lots of info)
-w<n> = override default 4-bit ADPCM width (2 <= n <= 5)
-x = exhaustive search (old behavior, very slow at depth)
-y = overwrite outfile if it exists
Web: Visit www.github.com/dbry/adpcm-xq for latest version and info
-
Unknown RIFF chunk types are correctly parsed on input files, but are not passed to the output file.
-
In some situations, at high lookahead levels, the operation can get very slow or even seem to be stuck, however this will happen at much higher lookahead depths than before. The default level 3 should always be fine and then the user can simply try increasing levels until the time becomes untenable. The new quantization noise option (-n) can be used to determine if higher levels are providing improvement (lower numbers are better). Note that the flat noise option (-f) will provide the lowest measured noise, but the default dynamic noise shaping may provide less audible noise.
-
Pipes are not yet supported.