10 · Reference

Benchmarks

Numbers with the method next to them, including the rows where the library is slower.

Method

Every measurement runs in a separate process: peak RSS is process-wide, so timing and memory cannot share one.
Median and p95 over repeated runs; the corpus is generated by the harness so it is reproducible.
Correctness is checked in the same run: decoded PCM is compared with libsndfile.
Hardware: Apple M5 Pro, macOS 26.6.2, Python 3.14.7, release build with lto = "fat" and opt-level = 3.

Scoreboard

Axis	sound_rs	Competitor	Ratio
Import, cold process	1.1 ms	31.8 ms (soundfile)	29×
Metadata, 60 s file	1.2 ms	31.7 ms	26×
Decode WAV	131 ms	155 ms	1.2×
Decode FLAC	265 ms	465 ms	1.8×
Decode MP3	247 ms	272 ms	1.1×
Decode AAC	361 ms	not supported	—
Time to first sound, 10 min	8.8 ms	492 ms	56×
Peak RSS, 10 min file	43 MB	255 MB	6×
Playback overhead, 3 s tone	32 ms	148 ms	4.6×
Batch decode, 2 FLAC files	272 ms	448 ms	1.6×

Decode detail

60 seconds of stereo 44.1 kHz, median and p95 over seven runs.

Format	sound_rs	libsndfile	PCM check
WAV 16-bit	131 / 160 ms	155 ms	identical
FLAC	265 / 267 ms	465 ms	identical
MP3	247 / 249 ms	272 ms	1.6·10⁻⁵, 47 samples offset
M4A (AAC)	361 / 365 ms	unsupported	8·10⁻⁸ vs ffmpeg

Playback and memory

	First sound	Peak RSS	Overhead
sound_rs Player, 10 min stream	8.8 ms	43 MB	+32 ms
soundfile + sounddevice	492 ms	255 MB	+148 ms

The playback overhead row is a 3-second tone: wall clock minus the track duration. The player number uses a stream that is already open; the one-shot play_array pays the device setup and lands around +180 ms.

Reproduce it

python bench/run.py                 # sound_rs only
python bench/run.py --compare       # + soundfile / sounddevice
python bench/run.py --reps 9

The harness generates its corpus (3 s and 10 min, WAV plus MP3/FLAC/M4A when ffmpeg is available), runs each measurement in a subprocess, and prints median, p95 and peak RSS. A full run takes about three minutes on the machine above.

Where it loses

MP3 alignment: 47 samples (~1 ms) earlier than libsndfile and ffmpeg. Length is exact; values match after alignment.
WAV decode: a memcpy race with libsndfile, 1.2× is all there is.
AAC decode is slower than the other formats — a newer decoder with less tuning — though libsndfile cannot decode it at all.
Very small files: for a 3-second tone the fixed device setup dominates everything, and all libraries land within a few tens of milliseconds of each other.

← API reference Troubleshooting →