syamaner/coffee-first-crack-audio

Coffee First Crack Audio Dataset

Audio dataset for training coffee first crack detection models. Contains 10-second WAV chunks from coffee roasting recordings, labelled as first_crack (popping/cracking sounds) or no_first_crack (background roast noise).

Model: syamaner/coffee-first-crack-detection Source code: github.com/syamaner/coffee-first-crack-detection Project home & essays: sertan.com/work

How this was built:

• Part 1 — The Architecture & The Agent — Spec-Driven ML Development with Warp/Oz

Original prototype:

• Part 1 — Training a Neural Network to Detect Coffee First Crack from Audio

Last updated: 12 Jul 2026 (baseline_v5, 303-sample test set).

Dataset Summary

• 1,435 total chunks (fixed 10-second sliding windows, no overlap)
• 21 source recordings from 4 microphone configurations, 4 coffee origins
• Recording-level split (no data leakage between splits)
• ~16% first_crack / ~84% no_first_crack — realistic class imbalance

Split	first_crack	no_first_crack	Total	Recordings
Train	136	786	922	14
Val	45	165	210	3
Test	42	261	303	4

Annotation Approach

Each source recording was annotated in Label Studio with a single first_crack region spanning from the first audible pop to the end of consistent cracking. The chunk_audio.py script then slid fixed 10-second windows across each recording and labelled each window based on overlap (>=50% threshold) with annotated first_crack regions.

This approach replaces the prototype method of manually annotating 20-30 small regions per file, producing consistent real-audio training chunks that match what the model sees during inference.

Features

Feature	Type	Description
audio	Audio (16kHz)	10-second mono WAV chunk
label	string	first_crack or no_first_crack
label_id	int	1 = first_crack, 0 = no_first_crack
microphone	string	e.g. mic-1-original, mic-2-new, mic-1-new, mic-2-new
coffee_origin	string	e.g. brazil, costarica-hermosa, brazil-santos, panama-hortigal-estate

Source Recordings

Current baseline (baseline_v5, 21 recordings): 9 legacy mic-1 + 6 amplified mic-2 + 3 mic-1-panama + 3 mic-2-panama. Reconciled against data/splits/split_report.md on 12 Jul 2026.

Mic	Origin	Recordings	Notes
mic-1-original	costarica-hermosa	5	Legacy recordings from prototype (4 hp-a + 1 .alog)
mic-1-original	brazil	4	Legacy recordings from prototype (.alog)
mic-2-new	brazil	4	Single-mic recordings (Feb 2026), gain-amplified
mic-2-new	brazil-santos	2	Single-mic recordings (Apr 2026), gain-amplified
mic-1-new (FIFINE)	panama-hortigal-estate	3	Multi-mic paired recordings
mic-2-new (ATR2100x)	panama-hortigal-estate	3	Multi-mic paired recordings

Usage

from datasets import load_dataset

ds = load_dataset("syamaner/coffee-first-crack-audio")
print(ds)
# DatasetDict({
#     train: Dataset({features: [audio, label, ...], num_rows: 922})
#     val: Dataset({features: [audio, label, ...], num_rows: 210})
#     test: Dataset({features: [audio, label, ...], num_rows: 303})
# })

# Access a sample
sample = ds["train"][0]
print(sample["label"], sample["microphone"], sample["coffee_origin"])

Citation

@misc{yamaner2026coffeefc,
  author = {Yamaner, Sertan},
  title  = {Coffee First Crack Audio Dataset},
  year   = {2026},
  url    = {https://huggingface.co/datasets/syamaner/coffee-first-crack-audio}
}