quinnlue/audioset-opus-fbanks

AudioSet Opus — INT8 log-mel filterbanks

Precomputed Kaldi log-mel filterbank features for AudioSet, quantized to INT8. Derived from danjacobellis/audioset_opus_24kbps and danjacobellis/audioset_opus_24kbps_balanced, which were already deduplicated by content hash.

The point of this dataset is to remove audio decoding and filterbank computation from the training loop. In a masked-autoencoder training step at 64×1008 geometry, the fbank front end costs ~41% of wall-clock at 1M parameters, ~15% at 15M and ~3.6% at 100M.

Feature specification

front end	Kaldi fbank (htk_compat=True, use_energy=False, window_type="hanning", dither=0.0)
sample rate	16,000 Hz (mono; source resampled with soxr where needed)
frame shift / length	10 ms / 25 ms
mel bins	64
frames	1008 (10 s of audio → 998 real frames, zero-padded to 1008)
stored dtype	int8, shape (1008, 64), C-order, 64,512 bytes/clip
stored quantity	raw log-mel — no mean/std normalization applied

Waveforms are DC-offset removed (per-clip mean subtraction) before the filterbank, matching the reference implementation.

Columns

column	type	meaning
path	string	source path, joinable against the upstream datasets
label	list<int64>	AudioSet class indices
fbank	binary	64,512 int8 bytes, reshape to (1008, 64)
scale	float32	per-clip dequantization scale
offset	float32	per-clip dequantization offset
n_frames	int16	real frames before zero-padding

Dequantization

Quantization is a per-clip affine map: each clip gets its own scale and offset from its own min/max, which uses the full 8-bit range for every clip.

import numpy as np

def dequantize(row):
    q = np.frombuffer(row["fbank"], dtype=np.int8).reshape(1008, 64)
    return (q.astype(np.float32) + 128.0) * row["scale"] + row["offset"]

Measured fidelity against float32, over 2,048 clips:

scheme	RMS error	max error	SNR
per-clip affine (used here)	0.0217	0.0494	45.7 dB
global affine (not used)	0.0260	3.0748	44.1 dB

RMS error is 0.0024 on the EAT-normalized scale, where the signal standard deviation is ~0.5 — roughly 0.5% of signal.

To apply the EAT/AudioMAE normalization used by most consumers:

features = (dequantize(row) - (-4.268)) / (4.569 * 2.0)

Splits

split	rows	source	AudioSet subset
train	1,912,024	audioset_opus_24kbps	unbalanced_train
validation	20,550	audioset_opus_24kbps_balanced (train)	balanced_train
test	18,886	audioset_opus_24kbps_balanced (validation)	eval_segments

test is the canonical AudioSet evaluation set. validation is balanced_train, held out here so it can serve as a fixed, mixture-independent scoring set. No clip appears in more than one split.

Provenance

Generated with ams.model.fbank.KaldiFbank from audio-mixture-scaling at 64 mel bins / 1008 frames. Decoding used torchcodec AudioDecoder with soxr resampling. Zero decode failures across all 1,951,460 clips (1,912,024 + 20,550 + 18,886).

License and attribution

AudioSet is released by Google under CC BY 4.0; the underlying audio comes from YouTube and remains subject to its original terms. These are lossy derived features (64-bin log-mel at 10 ms, INT8) from which the source audio cannot be reconstructed intelligibly. Please cite AudioSet and the upstream Opus datasets.