Auracle

Can a machine hear what a person hears?

Machine listening built from scratch in numpy rather than called from a library. Not a course. One question, and whatever I have to learn to answer it.

Auracle = aura + oracle, landing on auricle, the outer ear.

---

The structure

Every entry follows the same three beats:

1. The ear does this. 2. The machine does this. 3. Here is where the machine fails.

The third beat is the main goal. Every learning log on the internet has resources and code. However, not all of them explain why it fails from a humanistic perspective.

The labs

	Title	Status
01	Air pressure becomes a number	done
02	Frequency becomes pitch	done
03	Amplitude becomes loudness	done
04	The perceptual frequency axis	done
05	Timbre, or why a violin is not a flute	done
06	Pitch that isn't there	done
07	The beat is not in the signal	done
08	The cocktail party problem	done
09	Source separation, the learned answer	done
10	Tonality and key
11	Repetition, structure, and expectation
12	DSP in C++
13	Auto-tagging: from features to labels
14	Deep machine listening
15	The original question
16	Did any of it beat chance?
17	Ship

Full plan: docs/plans/2026-08-16-auracle-design.md Logs: docs/log/

Results

Filled in near the end, after every claim in this repo is tested against a shuffled-label null. Anything that does not survive gets retracted here, visibly.

Claim	Lab	Beat chance?
pending

Setup

uv venv
source .venv/bin/activate
uv pip install -r requirements.txt

Running a lab

source .venv/bin/activate
python labs/day-01-sampling/aliasing.py

Sources

Built from primary sources: papers, library documentation, and the canonical texts. The backbone reference is Meinard Müller, Fundamentals of Music Processing (FMP notebooks, free and executable). Citations live in each log entry.

License

MIT for the code. Written notes are the author's.