quinnlue/audioset_melspec_64_int8

AudioSet 64-bin INT8 log-mel spectrograms

Precomputed, normalized 1024×64 log-mel inputs derived from danjacobellis/audioset_opus_24kbps (train), plus the train and validation splits of danjacobellis/audioset_opus_24kbps_balanced.

Splits

Split	Source	Rows	Shards
train	Full AudioSet Opus train	1,912,024	96
balanced_train	Balanced AudioSet Opus train	20,550	2
validation	Balanced AudioSet Opus validation	18,886	2

The same validation-derived per-mel calibration is used for every split. balanced_train reconstruction measured 0.00270826 RMSE and 0.00224933 MAE, with 0.00000000% low-side and 0.00891425% high-side clipping.

Each melspec value is exactly 65,536 bytes in time-major C order. Decode with:

import json
import numpy as np

cal = json.load(open("calibration.json"))
scale = np.asarray(cal["scale"], dtype=np.float32)
offset = np.asarray(cal["offset"], dtype=np.float32)
q = np.frombuffer(row["melspec"], dtype=np.uint8).reshape(1024, 64)
spec = q.astype(np.float32) * scale + offset

Quantization uses independently calibrated affine ranges for all 64 mel bins. Candidate clipping percentiles were selected on the validation split by minimum reconstruction MSE. Calibration RMSE was 0.00263567; low/high clipping fractions were 0.00000000% and 0.00737877%. See dataset_manifest.json and calibration.json for full preprocessing and quantization metadata.