A VOX ADPCM (a.k.a. Dialogic ADPCM or OKI ADPCM) decoder for the NES. It may or may not be adaptable to other 6502-based systems. The code is written for ca65, but should be easily adaptable to other assemblers.
To avoid confusion, in this document, the term "sound clip" will be to used to refer to, well, sound clips. The term "sample" will refer not to a sound clip, but rather the samples sent to the DAC.
The engine has three primary modes, numbered 0 through 2, which we might call "slow", "faster", and "fastest". Each level of speed adds more performance at the expense of size (roughly 400 bytes per level). Maximum playback rates are as follows:
- Slow mode: 6351 Hz
- Faster mode: 8470 Hz
- Fastest mode: 8948 Hz
The mode can be changed by changing the SPEED variable in adpcm.asm. Note that sound clips can be big enough that you might not want much more than 6000 Hz anyway, in which case you may as well use slow mode.
In all modes, the sampling rate can be reduced by inserting delays between samples, allowing you to use smaller sound clips to save space. The playback engine is not perfectly timed, especially in the slower modes, and so the playback rate may be slightly unsteady. For practical purposes, however, it is good enough.
To save time, the program does not bother checking if the waveform clips. If you're getting clicks or pops, try reducing volume by 1 or 2 dB.
This program hogs the CPU; you cannot do anything while a sample is playing if you're using the maximum playback rate. If you're playing at, say, 4000 Hz, you get about 166 cycles of time between samples to do whatever it is.
VOX ADPCM is admittedly not the ideal format for playing on 8-bit systems with 7-bit audio. Each sample is actually decoded as 12-bit even though only the 7 most significant bits are heard. This means the program does a fair bit of 16-bit math. A codec that requires only 8-bit math to decode could run significantly faster and so yield a higher maximum sampling rate. But I wanted an ADPCM solution and I wanted it now, and it seemed easier to adapt readily available VOX ADPCM code in C than to find another solution.
Do note that the code doesn't bother to reset $4011 to zero after playing a sample. This means your triangle and noise channels might be a little quiet until you poke a new value into $4011.
Converting your samples to VOX ADPCM is easy! For our example, we're going to use Audacity, but other tools can be used.
First, load your sample in Audacity. Then, in the bottom-left corner of your screen, click the "Project Rate (Hz)" and type in the playback rate of the engine. For instance, if you're using the maximum playback rate of the engine's slow mode, type 6351.
Next, go to File->Export..., then before you save, set "Save as type" to "Other uncompressed files", then click the Options... button in the dialog box. Choose "RAW (header-less)" for the header and "VOX ADPCM" for the encoding. Now save your file and voila!
Sometimes this may produce a dialog box claiming an error of "No error". Don't worry; it probably still worked fine.
This is not necessary unless you modify the engine in some way.
Use FCEUX and break on writes to $4011. Ignore the first write and click "Run" to continue until it hits the watchpoint again. Somewhere in the middle of the right pane it should say "CPU cycles". Ignore the big number, but pay attention to the number in parentheses. It should say something like "(+288)". That's how many cycles it ran since the last break. Record several of these numbers in a row and average them. Then you have your average sampling rate in cycles.
To get a figure in Hz, use the formula 1,789,773 / cycles. (The NES runs at 1.789773 MHz.) For instance, if the average sampling rate is 281.8 cycles per sample, then the playback rate will be about 6351 Hz.
Suppose 6351 Hz sound clips take up too much space and you decide you'd be content with 4000 Hz sound clips. (Probably no good for human voices, but it may work fine for other things.) Then you can use the formula 1,789,773 / Hz to get the number of cycles you need per sample. 1,789,773 / 4000 = 447. Since the loop runs at 281 cycles per sample, you need to insert a delay of 447 - 281 = 166 clocks after each sample. An example loop is already in place in the code and will be used if you change SRATE_4000 to 1 at the top.