Another Audio Pedal Project (aapp) > phase 0: DSP

April 14, 2026

Why DSP

Ever since I started using my first DAW, I’ve been curious about the computing power needed to run all those effects and plugins on the sound you’re working with. I thought it would be interesting to dig a little deeper and understand the concepts behind these effects and plugins. So, what better project than an audio pipeline that mimics a pedal to see how everything comes together? I still need to decide the direction of the project but the main idea is to develop the backend and then maybe adding some UI elements.

phase 0: “Lets just start at zero; Level Zero”.

Getting mic input to come out of the speakers.

what I was trying to do

Phase 0 of my audio effects pedal project(The code is not public yet, but I will like to define a more long term plan first):

• plug mic in, hear audio out, no effects. Just prove the plumbing works before I start building the actual thing.

what actually happened

Installed PyAudio, wrote a scaffolding with the help of Claude and I was up and running the classic callback loop.

Running the python file and the output blessed my ears with a audio effect for crackling. Not quite what I was expecting but a start.

def callback(in_data, frame_count, time_info, status):
    samples = np.frombuffer(in_data, dtype=np.float32)
    return (samples.tobytes(), pyaudio.paContinue)

Looked fine to me. Wasn’t fine.

After some finicking I realized the issue: np.frombuffer returns a read-only array. The moment a downstream effect tries to write to it instant xrun(fancy word for de-syncronization between software and hardware), instant crackle. The fix is one word:

samples = np.frombuffer(in_data, dtype=np.float32).copy()

.copy(). That’s it of course.

The second issue was buffer size. I started with CHUNK = 512 because I recalled using this to export audio files from my DAW. Turns out 256 is the sweet spot on my machine small enough for sub-6ms latency (my current laptop is not of the fastest).

CHUNK = 256   # ~5.8ms @ 44100 Hz  tune this per machine
RATE  = 44100

what I learned

Two things stuck with me:

1. Never touch in_data directly always .copy() before processing.

2. Buffer size is a tradeoff, not a setting, smaller = lower latency but more CPU pressure and more xrun risk. 256 works for me, might not work for you. Profiling is a matter with which I will like to deal later on.

what’s next

Phase 1 — biquad EQ. I’ve already downloaded the Audio EQ Cookbook PDF. Looks intimidating :’)

Part of my aapp project.