Benchmarks

Wall time per feature on one 10 s clip at 22.05 kHz (a 220 Hz tone plus white noise, float32), with the parameters of the bundled config.json. Produced by

python benchmarks/bench_features.py --repeat 7 --librosa --torch-batch 16 --markdown

Each number is the median of 7 timed calls after one untimed warm-up; the IQR column is the interquartile range of those 7 calls in ms. AFX's STFT cache is cleared before every AFX call, so each feature pays for its own transform. librosa runs with the same n_fft, hop_length and center and with its default (disabled) cache. The torch column is one AFXFeatureModule forward pass under torch.no_grad() on a (16, 220500) float32 batch, divided by 16.

Machine: a 2 vCPU cloud VM, Intel Xeon Processor @ 2.10 GHz (model string as reported by lscpu, 1 thread per core, 4 MiB L2 per core), Linux 6.18, 7 GB RAM. Python 3.13.16, afx-audio 0.3.0, NumPy 2.5.3, SciPy 1.18.1, librosa 1.0.0, torch 2.14.1+cpu with 2 intra-op threads.

Feature	AFX ms	IQR	librosa ms	librosa function	torch ms per clip (batch of 16)
zcr	2.14	0.31	5.25	feature.zero_crossing_rate	5.74
variance	0.47	0.13
entropy	16.60	3.46
crest_factor	1.84	0.06
kurtosis	0.76	0.04
autocorr_variance	9.09	0.66
mobility	1.10	0.02
complexity	1.36	0.09
sample_entropy	267.52	4.15
energy_ratio	0.33	0.06
short_time_energy	0.36	0.03
rms_energy	0.36	0.02			0.08
spectral_centroid	8.27	0.19	8.71	feature.spectral_centroid	5.88
spectral_bandwidth	10.41	1.76	12.44	feature.spectral_bandwidth	6.86
spectral_rolloff	9.57	0.36	8.91	feature.spectral_rolloff	7.88
spectral_contrast	10.94	0.81	8.67	feature.spectral_contrast	18.71
spectral_entropy	10.24	0.38			7.35
spectral_flatness	9.39	0.44	6.74	feature.spectral_flatness	6.54
spectral_flux	11.23	0.34			5.68
spectral_skewness	12.36	0.54			10.26
spectral_slope	9.16	0.17			7.38
spectral_kurtosis	10.84	0.61			16.96
spectral_crest	8.90	0.78			6.85
spectral_decrease	9.32	0.40			5.91
melspectrogram	9.75	0.43	8.26	feature.melspectrogram	5.85
mfcc	9.98	0.62	8.48	feature.mfcc	6.14
mfcc_delta	10.59	0.77	10.10	feature.delta(mfcc)	6.28
mfcc_delta_delta	10.97	0.72	9.16	feature.delta(mfcc, order=2)	7.18
chroma_stft	9.68	0.86	7.21	feature.chroma_stft(tuning=0)	5.53
chroma_cqt	9.72	0.55	73.64	feature.chroma_cqt *
tonnetz	9.60	1.02	7.45	feature.tonnetz(chroma=chroma_stft)
onset_strength	9.98	0.50	8.25	onset.onset_strength
cqt	9.46	0.34	21.82	cqt *
pitch	19.81	0.54	22.98	yin *
thd	3.71	0.18
hnr	18.05	0.44
lpc	11.93	0.48
formants	46.31	2.99
gfcc	9.69	0.48			6.36
total	621.81				136.18 (one module)

extract_all_features on the same config, sharing one STFT per framing: 565.46 ms

Reading the table

Rows marked * are not the same computation. AFX's cqt and chroma_cqt weight STFT bins with log-spaced triangles, so they cost one STFT; librosa computes a true constant-Q transform with octave-wise filtering and resampling, which is 2 to 8 times slower here and has a time resolution that varies with frequency. pitch and librosa.yin implement the same estimator but differ in how unvoiced frames are reported (see librosa equivalence).

For the rows that are numerically equal to librosa, the two libraries are within about 40% of each other (zcr aside), because both are dominated by one real FFT over about 430 frames of 2048 samples. AFX also memoizes the mel and gammatone filterbanks across calls; librosa rebuilds its mel filterbank on every call.

sample_entropy is over 40% of the total. Its cost grows with the square of the frame length in the worst case (Richman and Moorman, 2000), and the bundled config computes it on frames of 1024 samples. Drop it, or use shorter frames, when throughput matters.

extract_all_features reuses one STFT for features with the same framing, centering and padding mode. With the bundled config that saves about 55 ms of the 620 ms sum; the rest of each spectral feature's time is the magnitude, power and per-feature reductions, which are not shared.

On this two-core machine the torch path runs most spectral features at the same speed per clip as the NumPy path or up to 2 times faster (spectral_flux 5.7 ms against 11.2 ms) by batching the FFT and matrix products. spectral_contrast (a sort per octave band), spectral_kurtosis and zcr are slower in torch. Expect the gap to change with core count and on a GPU; rerun the command above on the target machine.

Variance

On a shared VM, single runs move by 10 to 30%: runs of the command above on the same machine gave 622 to 678 ms for the total. The IQR column shows the spread within one run. Compare libraries within one run of the script, not across runs or machines.