| HN Mirror

blip_buffer and blip_buf are fairly decent tools for generating an output-rate signal from a high-rate signal (though I use my fork of blip_buffer at https://gitlab.com/exotracker/blip-buffer-exo and/or https://github.com/Dn-Programming-Core-Management/Dn-FamiTra...). When emulating sound chips, the blip buffer adds and subtracts bandlimited sinc impulses from an output-rate delta array, and when exporting resampled audio, it performs a running sum (transforming the impulses into steps) of the delta array and high-passes the running value when writing to the output audio array. The advantage is that you can pick very high sampling rates (like 1.79 MHz) but only burn CPU cycles each time the output level changes (unlike conventional resamplers which are O(input rate + output rate)). Unfortunately blip_buffer has a relatively low SNR of ~50 dB of aliasing rejection even at its widest impulse kernel (as measured by https://gitlab.com/exotracker/exotracker-cpp/-/tree/rewrite-..., though possibly blip_buf is better), and only performs an approximate highpass using integer bitshifts. A usage example is at https://github.com/Dn-Programming-Core-Management/Dn-FamiTra....

Alternatively you can generate a high-rate signal and feed it into a conventional resampler to produce a 44.1/48/96 KHz output. I found that libsamplerate (https://libsndfile.github.io/libsamplerate/)'s medium preset produces audibly transparent output at 44.1 KHz and above, and should have acceptable latency on the order of 1ms (I didn't verify but you could first flush out the startup edge effect with silence, pop all output, then push an impulse followed with silence until the central peak emerges from the output). This has minimal CPU usage for a single stereo 128 KHz input stream (like in exotracker and chipsynth SFC), but I don't know if it burns excessive CPU with 1.79 MHz input.

----

My baseline expectation for production-quality emulators is to generate sound without aliasing, but the gold standard is to properly emulate the audio path as found on hardware, by feeding schematics through SPICE and/or pole-zero math to create an analytical representation of the filters, then verifying them against MDFourier tests (hardware recordings of broad-spectrum sound played by the console). Few emulators attempt to do this; according to https://bel.fi/alankila/modguide/interpolate.txt, UADE (an Amiga emulator) gets this right using a variation of the Blip_Buffer approach with longer precomputed(?) impulse responses specialized for Amiga filtering. Several chiptune tools properly model hardware filters, including the chipsynth family of audio VSTs (commercial); Dn-FamiTracker (an open-source NES composer) emulates FDS lowpass properly without aliasing, but only loosely approximates 2A03 lowpass and global highpass using blip_buffer's configurable filtering (impulse/step visualizer at https://gitlab.com/exotracker/exotracker-cpp/-/blob/rewrite-...).

If you choose to model a hardware filter using IIR filters (mathematical arithmetic based off a hardware model) instead of a large precomputed impulse response (like interpolate.txt and UADE), you'll get more accurate results if you generate audio at a high internal sampling rate, IIR-filter the audio at this high rate (ensuring the filter cutoff is well below Nyquist or half the sampling rate), then feed it into a resampler. If you use Blip_Buffer to generate 44.1 or 48 KHz directly like blip_buffer, and apply a filter with cutoff above 10 KHz or so, high frequencies will not be filtered accurately.

One interesting idea (combining blip_buffer's efficiency at handling sparse signals, and the accurate treble filtering enabled by a high intermediate filtering frequency) is running a blip_buffer-like system (with no highpass but a ~20 KHz lowpass) to downsample from a high internal rate to a fixed 128 KHz (for fixed filtering) or twice the audio rate (for efficient rational-factor downsampling), then performing hardware filtering there before downsampling using a resampler. The downside is that this stacks the latency and artifacts of both Blip_Buffer and the resampler, but if you make Blip_Buffer generate mostly-lowpassed audio and avoid generating nonlinear harmonics in filtering, you can use a faster second resampler that assumes its input is mostly lowpassed (using a narrower sinc kernel).